Русский
EnglishCryptocurrencies:
8197 /
Markets: 113117
Market Cap: $ 2 080 075 612 221 /
24h Vol: $ 67 700 548 334 /
BTC
Dominance: 58.0939815373%
Новости
[Перевод] Учебник — всё, что вам нужно
Немного вызывающее название статьи отсылает к известной работе Внимание - всё, что вам нужно. На этот раз речь пойдет о качестве данных, на которых обучают LLM. Оказывается, качественный учебник (как концентрат знаний в любой сфере) в разы сокращает потребность и в памяти, и в мощности GPU, и в деньгах инвесторов...
Эта статья - о качестве данных при обучении/файнтюнинге LLM. Классическая работа о количественных данных позволит лучше понять, на что влияет качество датасета: Законы масштабирования нейронных языковых моделей
2 октября 2023
Аннотация
Мы представляем phi-1, новую большую языковую модель для генерации кода, которая значительно меньше по размеру по сравнению с конкурирующими моделями: phi-1 — это модель на основе архитектуры трансформера с 1,3 миллиардами параметров, обученная в течение 4 дней на 8 A100, с использованием выборки данных "качество-как-в-учебнике" из интернета (6 миллиардов токенов) а также синтетически сгенерированных с помощью GPT-3.5 учебников и упражнений (1 миллиард токенов). Несмотря на такой небольшой масштаб, phi-1 достигает точности pass@1 50,6% на HumanEval и 55,5% на MBPP. Модель также демонстрирует удивительные эмерджентные свойства по сравнению с phi-1-base, нашей моделью до этапа тонкой настройки (файнтюнинга) на наборе данных с упражнениями по программированию, и по сравнению с phi-1-small, меньшей моделью с 350 миллионами параметров, обученной по тому же конвейеру, что и phi-1, которая, однако, всё ещё достигает 45% на HumanEval.
1. Введение
Искусство обучения больших нейронных сетей показало невероятный прогресс за последнее десятилетие, особенно после открытия архитектуры трансформера [*], однако научные знания, стоящие за этим успехом, пока остаются ограниченными. Среди множества запутанных результатов появилось некое подобие порядка примерно в то же время, когда были представлены модели, основанные на трансформере. Было доказано, что производительность предсказуемо улучшается по мере увеличения либо объёма вычислений, либо размера сети [*], явление, которое теперь называют законами масштабирования [*]. Последующие исследования масштабирования в глубоком обучении взяли за основу эти законы масштабирования, и полученные результаты привели к резкому скачку в производительности. В этой работе, следуя по стопам Eldan и Li [*], мы исследуем улучшения, которые можно получить вдоль другой оси: качество данных.
Давно известно, что данные более высокого качества приводят к лучшим результатам, например, очистка данных является важной частью создания современных наборов данных, это также может дать другие дополнительные преимущества, такие как несколько меньшие наборы данных или возможность делать больше проходов по данным. Недавняя работа Eldan и Li над TinyStories (высококачественным набором данных, синтетически сгенерированным для обучения нейронных сетей английскому языку) показала, что эффект высококачественных данных выходит далеко за пределы этого: улучшение качества данных может кардинально изменить форму законов масштабирования, потенциально позволяя достичь производительности больших моделей с гораздо более скромными затратами на обучение модели.
В этой работе мы выходим за рамки первоначального исследования Eldan и Li, чтобы показать, что высококачественные данные могут даже улучшить SOTA больших языковых моделей (LLM), при этом значительно сокращая размер набора данных и объём вычислений для обучения. Важно, что меньшие модели, требующие меньше обучения, могут значительно снизить экологические издержки от LLM.
Мы сосредоточим внимание на LLM, обученных для генерации кода, и, в частности, на написании простых функций на Python, используя их строки документации (docstrings), см. [*]. Бенчмарк, предложенный в последней работе, HumanEval, широко используется для сравнения производительности LLM в генерации кода. Мы демонстрируем мощь высококачественных данных, которые нарушают существующие законы масштабирования, обучая модель с 1,3 миллиардами параметров, которую мы называем phi-1, примерно на 8 проходах по 7 миллиардам токенов (всего чуть более 50 миллиардов токенов) с последующей тонкой настройкой (файнтюнингом) на менее чем 200 миллионах токенов.
Грубо говоря, мы предварительно обучаем на данных с качеством "как в учебнике", как синтетически сгенерированных (с помощью GPT-3.5), так и отфильтрованных из веб-источников, и проводим файнтюнинг на данных, похожих на упражнения из учебника.
Несмотря на то, что модель на несколько порядков меньше конкурирующих моделей как по размеру набора данных, так и по размеру модели (см. Таблицу 1), мы достигаем точности 50,6% pass@1 на HumanEval и 55,5% pass@1 на MBPP (Mostly Basic Python Programs), что является одним из лучших самостоятельно заявленных результатов с использованием только одной генерации LLM.
Таблица 1: Мы используем данные отчетов, если они доступны. Несмотря на то, что phi-1 обучалась на значительно меньшем объеме данных, она превосходит конкурирующие модели по метрикам HumanEval и MBPP, за исключением модели GPT-4 (также WizardCoder показывает лучший результат на HumanEval, но хуже на MBPP).
Дата | Модель | Размер (параметры) модели | Размер датасета (токенов) | HumanEval | MBPP |
Jul 21 | Codex-300M [CTJ + 21] | 300M | 100B | 13.2% | N.A. |
Jun 23 | phi-1 | 1.3B | 7B | 50.6% | 55.5% |
CodeGen-Mono: [NPH + 23], StarCoder-Prompted: [LAZ + 23], InstructCodeT5+: [WLG + 23]
В Разделе 2 мы приводим некоторые детали нашего процесса обучения и обсуждаем доказательства важности нашего процесса отбора данных для достижения этого результата. Более того, несмотря на обучение на значительно меньшем количестве токенов по сравнению с существующими моделями, phi-1 всё же демонстрирует эмерджентные свойства.
Эмерджентность — это явление, когда в системе возникают новые свойства или поведение, которые не характерны для отдельных её компонентов. Т.е. это проявления феномена, когда целое больше, чем сумма его частей - прим. переводчика.
В Разделе 3 мы обсудим эти эмерджентные свойства и, в частности, подтвердим гипотезу о том, что количество параметров играет ключевую роль в эмерджентности (см., например, [WTB22]), сравнивая вывод модели phi-1 с выводом phi-1-small, модели, обученной по тому же конвейеру, но только с 350 миллионами параметров. Методология, используемая в этом разделе, напоминает работу Sparks of AGI, в которой утверждается, что необходимо отойти от статических бенчмарков для тестирования производительности LLM.
Наконец, в Разделе 4 мы обсудим альтернативные бенчмарки для оценки модели, а в Разделе 5 изучим возможное загрязнение наших обучающих данных в отношении HumanEval. Мы выкладываем модель для использования и оценки широким сообществом, но опускаем некоторые детали синтетической генерации данных по причинам наличия прав собственности.
Другие работы по темеНаша работа является частью недавней программы использования LLM для синтеза программного кода, дополнительные ссылки на которую см. в [CTJ+21, NPH+22]. Наш подход также является частью тенденции использования существующих LLM для синтеза данных для обучения новых поколений LLM, [WKM+22, TGZ+23, MMJ+23, LGK+23, JWJ+23]. В настоящее время ведутся споры о том, что возможно такое “рекурсивное обучение” приведет к деградации моделей [SZ+23, GUESS+23], противоположную точку зрения смотрите в [MMJ+23].
Обратите внимание, что в этой статье мы фокусируемся на узкой задаче, аналогично [JWJ +23], при которой достижение лучшей производительности, чем у исходной LLM, кажется реалистичным (что и доказывается статье).
2. Детали обучения и важность высококачественных данных
Рисунок 2.1: Точность Pass@1 (%) на HumanEval. Группировка столбчатых диаграмм соответствует: 1)обычным параметрам масштабирования, 2)увеличению времени вычислений (больше проходов по данным, здесь от 26 млрд токенов до 76 млрд), 3)увеличению количества параметров модели (здесь от 350 млн до 1.3 млрд). Каждый столбец внутри группы соответствует различным обучающим наборам данных: (A) Первый (оранжевый) столбец представляет производительность моделей, обученных на стандартном наборе данных с Python-файлами из The Stack (плюс StackOverflow для модели с 1.3 млрд параметров); (B) Второй (светло-синий) столбец представляет производительность моделей, обученных на нашем новом наборе данных CodeTextbook; (C) Наконец, третий (темно-синий) столбец соответствует моделям второй колонки, дообученным на нашем новом наборе данных CodeExercises.
Для моделей с 1.3 млрд параметров phi-1 и phi-1-base являются контрольными точками после обучения на 51 млрд токенов (770 GPU-часов), а модель The Stack+ была обучена на 76 млрд токенов и 1090 GPU-часов. Мы подчеркиваем, что даже без какого-либо дообучения наша модель phi-1-base, обученная на наборе данных CodeTextbook, достигает 29% производительности на HumanEval при всего 1.3 млрд параметров. Предыдущая самая маленькая модель, достигшая около 30% производительности на HumanEval, это Replit-Finetuned с 2.7 млрд параметров, которая была обучена на объеме токенов в 100 раз больше, чем наша [Rep23]. Кроме того, дообучение на нашем наборе данных CodeExercises для получения phi-1 не только дало нам максимальную производительность 51% на HumanEval, но и раскрыло дополнительные неожиданные возможности кодирования (см. Раздел 3).
Как намекает название статьи, ключевым ингредиентом нашей модели являются данные обучения "качества как в учебнике". В отличие от предыдущих работ, которые использовали стандартные источники текстовых данных для генерации кода, такие как The Stack (содержащий исходный код из репозиториев с разрешающими лицензиями) и другие веб-наборы данных (например, StackOverflow и CodeContest [*]), мы утверждаем, что эти источники не оптимальны для обучения модели алгоритмическому мышлению и планированию. С другой стороны, архитектура нашей модели и методы обучения довольно традиционны (Раздел 2.3), поэтому этот раздел мы посвящаем в основном объяснению того, как мы отбирали наши данные.
Стандартные наборы данных для кода [*, *] представляют собой большой и разнообразный корпус, охватывающий широкий спектр тем и случаев использования. Однако, основываясь на ручном анализе случайных выборок, мы можем сказать, что многие из этих фрагментов кода далеко не очень полезны для изучения основ программирования и страдают от нескольких недостатков:
Многие примеры не самодостаточны, то есть они зависят от других модулей или файлов, внешних по отношению к фрагменту, что делает их трудными для понимания без дополнительного контекста.
Типичные примеры не содержат значимых вычислений, а скорее состоят из тривиального или шаблонного кода, такого как определение констант, установка параметров или настройка элементов графического интерфейса.
Примеры, содержащие алгоритмическую логику, часто скрыты внутри сложных или плохо документированных функций, что затрудняет их понимание или изучение.
Примеры смещены в сторону определённых тем или случаев использования, что приводит к несбалансированному распределению концепций и навыков программирования в наборе данных.
Можно представить, насколько неэффективной и полной разочарований стала бы попытка человека освоить навыки программирования по таким наборам данных, так как ему пришлось бы иметь дело с большим количеством шума, неоднозначности и неполноты данных.
Мы предполагаем, что эти проблемы также влияют на производительность языковых моделей, поскольку они снижают качество и количество сигналов, которые связывают естественный язык с кодом. Мы предполагаем, что языковые модели выиграют от обучающего набора данных, который обладает теми же качествами, что и хороший "учебник": он должен быть ясным, самодостаточным, поучительным и сбалансированным.
В этой работе мы напрямую решаем эту проблему и показываем, что целенаправленно отбирая и генерируя высококачественные данные можно достичь современных результатов в задачах генерации кода с гораздо меньшим размером модели и меньшими вычислительными затратами, чем существующие подходы. Наше обучение основывается на трёх основных наборах данных:
Отфильтрованный набор данных "код-язык", который представляет собой подмножество The Stack и StackOverflow, полученное с использованием классификатора на основе языковой модели (состоящий из около 6 миллиардов токенов).
Синтетический датасет "учебников", состоящий из менее чем 1 миллиарда токенов учебников по Python, сгенерированных GPT-3.5.
Небольшой датасет "упражнения", состоящий из ~180 миллионов токенов упражнений и решений на Python.
Мы подробно описываем эти наборы данных в следующих подразделах. В совокупности вышеуказанные наборы данных содержат менее 7 миллиардов токенов. Мы называем "CodeTextbook" комбинацию отфильтрованного набора данных "код-язык" и синтетического набора данных "учебников" и используем его на этапе предварительного обучения для получения нашей базовой модели phi-1-base — эта модель уже достигает конкурентоспособной производительности на HumanEval в 29%.
Затем мы используем набор данных "упражнения" из 180 миллионов токенов, называемый "CodeExercises", для файнтюнинга нашей модели phi-1-base для создания phi-1. Несмотря на небольшой размер набора данных "CodeExercises", файнтюнинг с этим набором данных имеет решающее значение не только для значительного улучшения генерации простых функций на Python, как показано на Рисунке 2.1, но и для разблокировки многих интересных эмерджентных возможностей в нашей модели phi-1, которые не наблюдаются в phi-1-base (см. Раздел 3).
2.1 Фильтрация существующих наборов данных с использованием классификатора на основе трансформера
Мы начинаем с общедоступных наборов данных на Python: используем подмножество Python из дедуплицированной версии The Stack и StackOverflow, которые вместе содержат более 35 миллионов файлов/примеров, всего более 35 миллиардов токенов.
Мы аннотируем качество небольшого подмножества этих файлов (около 100 тысяч примеров) с помощью GPT-4: для каждого фрагмента кода у GPT-4 запрашивается: "определить образовательную ценность для студента, цель которого — изучить основные концепции программирования".
Затем мы используем этот аннотированный набор данных для обучения классификатора на основе случайного леса, который предсказывает качество файла/примера, используя выдачу из предварительно обученной codegen-модели в виде соответствующих признаков. Мы отмечаем, что, в отличие от GPT-3.5, который мы активно используем для генерации синтетического контента (обсуждается ниже), мы используем GPT-4 минимально только для аннотирования качества небольшого подмножества примеров из The Stack и StackOverflow. Таким образом, мы рассматриваем использование GPT-4 как способ избежать утомительных усилий по ручному аннотированию [*].
Образовательная ценность, определенная фильтром / Высокая ценность (слева) / Низкая ценность (справа)
Код выше в формате python:# Высокая ценность
import torch
import torch.nn.functional as F
def normalize(x, axis=-1):
"""Performs L2-Norm."""
num = x
denom = torch.norm(x, 2, axis, keepdim=True)
.expand_as(x) + 1e-12
return num / denom
def euclidean_dist(x, y):
"""Computes Euclidean distance."""
m, n = x.size(0), y.size(0)
xx = torch.pow(x, 2).sum(1, keepdim=True).
expand(m, n)
yy = torch.pow(x, 2).sum(1, keepdim=True).
expand(m, m).t()
dist = xx + yy - 2 * torch.matmul(x, y.t())
dist = dist.clamp(min=1e-12).sqrt()
return dist
def cosine_dist(x, y):
"""Computes Cosine Distance."""
x = F.normalize(x, dim=1)
y = F.normalize(y, dim=1)
dist = 2 - 2 * torch.mm(x, y.t())
return dist# Низкая ценность
import re
import typing
...
class Default(object):
def __init__(self, vim: Nvim) -> None:
self._vim = vim
self._denite: typing.Optional[SyncParent] = None
self._selected_candidates: typing.List[int] = []
self._candidates: Candidates = []
self._cursor = 0
self._entire_len = 0
self._result: typing.List[typing.Any] = []
self._context: UserContext = {}
self._bufnr = -1
self._winid = -1
self._winrestcmd = ''
self._initialized = False
self._winheight = 0
self._winwidth = 0
self._winminheight = -1
self._is_multi = False
self._is_async = False
self._matched_pattern = ''
...Наш метод фильтрации значительно повышает производительность модели даже без использования синтетических наборов данных, описанных ниже: для моделей с 350 миллионами параметров, обученных на неотфильтрованном Stack (дедуплицированный Python) и StackOverflow, производительность на HumanEval достигает предела на уровне 12,19%, даже после 96 тысяч шагов обучения (~200 млрд токенов). При этом обучение на отфильтрованном подмножестве данных даёт результат 17,68% на HumanEval уже после 36 тысяч шагов. Мы дополнительно улучшили этот показатель до 20,12% (см. рисунок 2.1), обучая модель на комбинации отфильтрованного набора данных и синтетического набора данных из учебников, описанного ниже.
2.2 Создание синтетических наборов данных с "качеством учебника"
Одной из основных проблем при создании высококачественного набора данных для генерации кода является обеспечение того, чтобы примеры были разнообразными и не повторялись.
Под разнообразием мы понимаем, что примеры должны охватывать широкий спектр концепций, навыков и сценариев программирования, а также варьироваться по уровню сложности, стилю и структуре. Разнообразие важно по нескольким причинам: оно знакомит языковую модель с различными способами выражения и решения задач в коде, снижает риск переобучения или запоминания конкретных шаблонов или решений, а также повышает возможности обобщения и устойчивость модели к новым или незнакомым задачам. Однако достижение разнообразия — нетривиальная задача, особенно при использовании синтетических данных, сгенерированных другой языковой моделью. Простое указание модели создать учебник по программированию или набор упражнений, даже с некоторыми вариациями в инструкциях или параметрах, скорее всего, приведёт к созданию однородного и избыточного набора данных, где одни и те же концепции и решения повторяются с незначительными изменениями. Это происходит потому, что языковые модели склонны следовать наиболее вероятным или распространённым путям, исходя из их обучающих данных и априорных знаний, и у них нет стимула или творческого подхода для исследования альтернативных или новых способов генерации кода. Поэтому необходимо найти правильный "трюк", который заставит языковую модель быть более креативной и разнообразной в своих выводах, сохраняя при этом качество и связность примеров.
Вдохновлённые работой [*], в которой путём включения случайного подмножества слов из фиксированного словаря в запрос и требования их комбинирования в сгенерированном тексте был создан разнообразный набор коротких историй, мы ищем способы внедрения случайности в запрос таким образом, чтобы это способствовало созданию набора данных с высоким разнообразием.
Синтетический набор данных "учебников"
Этот набор данных состоит из менее чем 1 миллиарда токенов учебников по Python, сгенерированных GPT-3.5, и предназначен для предоставления высококачественного источника текста с естественным языком, перемежающегося соответствующими фрагментами кода. Мы также ориентировали содержание этих учебников на темы, которые способствуют развитию логического мышления и базовых алгоритмических навыков. Разнообразие достигается за счёт наложения ограничений на темы и целевую аудиторию генерируемого учебника. Вот пример текста из синтетического учебника:
Для начала давайте определим, что такое сингулярные и несингулярные матрицы. Матрица называется сингулярной, если её определитель равен нулю. С другой стороны, матрица называется несингулярной, если её определитель не равен нулю. Теперь давайте рассмотрим эти концепции на примерах.
Пример 1: Рассмотрим матрицу
A = np.array([[1, 2], [2, 4]]). Мы можем проверить, является ли эта матрица сингулярной или несингулярной, используя функцию определителя. Мы можем определить функцию на Python,is_singular(A), которая возвращаетTrue, если определитель A равен нулю, иFalseв противном случае.
import numpy as np
def is_singular(A):
det = np.linalg.det(A)
if det == 0:
return True
else:
return False
A = np.array([[1, 2], [2, 4]])
print(is_singular(A)) # TrueНабор данных "CodeExercises"
Это небольшой синтетический набор данных "упражнения", состоящий из менее чем 180 миллионов токенов упражнений и решений на Python. Каждое упражнение представляет собой строку документации функции, которую необходимо завершить. Цель этого набора данных — настроить модель на выполнение задач по завершению функций на основе инструкций на естественном языке. Этот набор данных также был сгенерирован GPT-3.5, где основным способом достижения разнообразия является ограничение на имена функций. Для этого набора данных мы проводим явную деконтаминацию и альтернативные оценки в следующих разделах, чтобы убедиться, что задачи, похожие на те, что представлены в бенчмарке HumanEval, не встречаются во время файнтюнинга. Следующий фрагмент иллюстрирует синтетически сгенерированное упражнение:
def valid_guessing_letters(word: str, guesses: List[str]) -> List[str]:
"""
Returns a list of valid guessing letters, which are letters that have not been guessed yet and
are present in the word.
Parameters:
word (str): The word to guess.
guesses (List[str]): A list of letters that have already been guessed.
Returns:
List[str]: A list of valid guessing letters.
"""
valid_letters = []
for letter in word:
if letter not in guesses and letter not in valid_letters:
valid_letters.append(letter)
return valid_letters2.3. Архитектура модели и обучение
Мы использовали трансформер с декодером [*], использующий реализацию FlashAttention для многоголового внимания (MHA) [*]. Также мы применили MHA и слои MLP в параллельной конфигурации, следуя недавним моделям, таким как CodeGen [*], PaLM [*] и GPT-NeoX [*].
Архитектура нашей модели phi-1 с 1.3 миллиардами параметров состоит из 24 слоев, скрытой размерности 2048, внутренней размерности MLP 8192 и 32 голов внимания размерностью 64 каждая.
Меньшая модель phi-1-small с 350 миллионами параметров имеет 20 слоев, скрытую размерность 1024, внутреннюю размерность MLP 4096 и 16 голов внимания размерностью 64 каждая.
Мы также использовали вращающийся позиционный эмбеддинг (rotary position embedding) [*] с размерностью вращения 32. Эти архитектурные решения были заимствованы из [*]. Мы использовали тот же токенизатор, что и в codegen-350M-mono [*]. Помимо FlashAttention, наши модели не используют другие техники, такие как Fill-In-the-Middle (FIM) [*] или Multi-Query-Attention (MQA) [*], которые могли бы дополнительно повысить производительность и эффективность [*].
Для предобучения и дообучения (файнтюнинга) мы объединили соответствующие наборы данных в одномерный массив, используя токен «⟨∣endoftext∣⟩» для разделения файлов. Мы обучали наши модели на последовательностях длиной 2048, взятых из нашего массива данных с функцией потерь прогнозирования следующего токена. Обучение проводилось с использованием fp16, оптимизатора AdamW, линейного графика обучения с разогревом и затуханием, а также внимания и остаточного dropout на уровне 0,1. Обучение выполнялось на 8 GPU Nvidia A100 с использованием deepspeed. Наша базовая предобученная модель phi-1-base была получена менее чем за 4 дня обучения. Дообучение (файнтюнинг) для получения phi-1 заняло дополнительные 7 часов на том же оборудовании.
Предобучение (Pretraining)
phi-1-base был обучен на датасете CodeTextbook (отфильтрованный корпус кода и синтетические учебники). Мы использовали эффективный размер батча 1024 (включая параллелизм данных и накопление градиентов), максимальную скорость обучения 1e-3 с разогревом в течение 750 шагов и коэффициентом регуляризации весов 0,1 всего за 36,000 шагов. Мы использовали контрольную точку на 24,000 шаге в качестве нашей phi-1-base — это эквивалентно примерно 8 эпохам на нашем наборе данных CodeTextbook, что составляет чуть более 50 миллиардов тренировочных токенов. Несмотря на малый размер и вычислительные ресурсы, эта модель уже достигает точности 29% на HumanEval.
Файнтюнинг
phi-1 получена путем файнтюнинга phi-1-base на наборе данных CodeExercises. Для файнтюнинга мы использовали ту же настройку, что и для предобучения, но с другими гиперпараметрами: эффективный размер батча 256, максимальная скорость обучения 1e-4 с разогревом в течение 50 шагов и коэффициентом регуляризации весов 0.01. Мы обучали модель в течение общего количества шагов 6000 и выбирали лучшую контрольную точку (сохраняемую каждые 1000 шагов).
3. Скачок возможностей модели после файнтюнинга на наборе CodeExercises
На Рисунке 2.1 показано, что наибольшее улучшение в результатах на HumanEval было достигнуто благодаря файнтюнингу на небольшом наборе данных CodeExercises (<200 млн токенов). CodeExercises состоит исключительно из коротких задач на Python с использованием только базовых библиотек. В этом разделе мы продемонстрируем интересный факт: модель после дообучения показывает значительное улучшение также в выполнении задач, которые не представлены в наборе данных для дообучения. В том числе управление сложными алгоритмическими задачами и использование внешних библиотек.
Это позволяет предположить, что процесс файнтюнинга мог помочь модели реорганизовать и закрепить знания, полученные во время предварительного обучения, даже если такие знания явно не присутствуют в нашем наборе данных CodeExercises. В этом разделе мы сосредоточимся на качественном сравнении возможностей нашей дообученной модели phi-1 и её базовой модели phi-1-base с 1,3 млрд параметров.
3.1 Файнтюнинг улучшает понимание модели
Используя простую функцию Python, которую мы создали самостоятельно, мы наблюдаем, что модель демонстрирует значительно более высокий уровень понимания и соответствия инструкциям после файнтюнинга. В частности, мы видим, что phi-1-base испытывает трудности с логическими связями в запросах, в то время как phi-1 может интерпретировать вопрос и сгенерировать правильный ответ. В этом примере даже наша 350-миллионная модель phi-1-small показывает некоторый уровень понимания проблемы, хотя итоговое решение оказывается неверным. Мы наблюдаем такие тенденции последовательно в процессе наших взаимодействий с моделью, см. другой пример в Приложении A.
Промпт: Алиса, Боб и Чарльз играют в игру. Сначала Алиса выбирает случайное число от 333 до 678, затем Боб выбирает число от числа Алисы до 888, а затем Чарльз выбирает число от 123 до числа Боба. Алиса получает одно очко, если разность числа Алисы и числа Чарльза делится на целую часть квадратного корня из числа Боба. В противном случае Боб получает одно очко. Смоделируйте очки Алисы и Боба за 30 итераций.
Код предыдущего примера:# phi-1
points = [0, 0]
for i in range(30):
alice_num = random.randint(333, 678)
bob_num = random.randint(alice_num, 888)
charles_num = random.randint(123, bob_num)
if (alice_num - charles_num) % int(bob_num ** 0.5) == 0:
points[0] += 1
else:
points[1] += 1
return f"Alice: {points[0]}, Bob: {points[1]}"# phi-1-base
self.points = 0
self.alice = 333
self.bob = 678
self.charles = 888
self.bob_points = 0
self.alice_points = 0
self.iterations = 30
self.alice_turn = True
self.bob_turn = True
self.alice_turn_points = 0
self.bob_turn_points = 0
self.alice_turn_points_history = []
self.bob_turn_points_history = []
self.alice_turn_points_history.append(self.alice_points)
self.bob_turn_points_history.append(self.bob_points)# phi-1-small
indices = []
import random
alice_points = 0
bob_points = 0
for i in range(30):
alice_num = random.randint(333, 678)
bob_num = random.randint(123, bob_num)
if (alice_num - bob_num) % math.floor(math.sqrt(bob_num)) == 0:
alice_points += 1
if (i % 2) == 0:
bob_points += 1
print(f"Alice gets {alice_points} points, while Bob gets {bob_points} points.")3.2. Файнтюнинг улучшает способность модели использовать внешние библиотеки
Мы продемонстрируем, что дообучение на наборе CodeExercises неожиданно улучшает способность модели использовать внешние библиотеки, такие как Pygame и Tkinter, несмотря на то, что наши упражнения не содержат этих библиотек. Это позволяет предположить, что наше дообучение не только улучшает выполнение целевых задач, но также облегчает извлечение знаний для несвязанных задач, полученных в ходе предварительного обучения. Для справки, на Рисунке 3.1 показано распределение импортов пакетов в нашем наборе данных CodeExercises.
Рисунок 3.1: Количество импортов среди ∼880 тыс. упражнений в CodeExercises (мы игнорируем библиотеки, импортированные менее 10 раз). Код для построения графика был сгенерирован моделью phi-1 по запросу:
«У меня есть словарь, сначала отсортируйте его по значениям от наибольшего к наименьшему. Затем постройте столбчатую диаграмму с помощью pyplot. Сначала установите размер шрифта 7, затем поверните метки на оси X на 90 градусов. Ось X — это ключи, ось Y — значения словаря. Используйте логарифмическую шкалу для оси Y. Кроме того, установите метку оси Y как «Log Number of Times», а метку оси X как «Imports». Установите dpi на 1000.»
«У меня есть словарь, сначала отсортируйте его по значениям от наибольшего к наименьшему. Затем постройте столбчатую диаграмму с помощью pyplot. Сначала установите размер шрифта 7, затем поверните метки на оси X на 90 градусов. Ось X — это ключи, ось Y — значения словаря. Используйте логарифмическую шкалу для оси Y. Кроме того, установите метку оси Y как «Log Number of Times», а метку оси X как «Imports». Установите dpi на 1000.»
Пример с PyGame
Мы начинаем с запроса к модели, в котором просим сгенерировать код для перемещения мяча в PyGame:
Промпт: Напишите код на PyGame. Есть мяч. На каждой итерации: 1. Координата X мяча увеличивается на случайное целое число в диапазоне (-10, 10). 2. Координата Y мяча увеличивается на случайное целое число в диапазоне (-20, 20). Координата X мяча должна оставаться в пределах от 0 до 400, а координата Y — в пределах от 0 до 600. Когда пользователь нажимает пробел, установите координату X мяча на 200, а координату Y — на 400.
Код примеров:# phi-1
...
screen = pygame.display.set_mode((400, 600))
x, y = 0, 0
while True:
for event in pygame.event.get():
...
elif event.type == pygame.KEYDOWN:
if event.key == pygame.K_SPACE:
x = 200
y = 400
x += random.randint(-10, 10)
y += random.randint(-20, 20)
if x < 0:
x = 0
elif x > 400:
x = 400
if y < 0:
y = 0
elif y > 600:
y = 600
screen.fill((255, 255, 255))
pygame.draw.circle(screen, (255, 0, 0), (x, y), 10)
pygame.display.update()# phi-1-base
...
ball_x = random.randint(-10, 10)
ball_y = random.randint(-20, 20)
ball_vx = random.randint(-5, 5)
ball_vy = random.randint(-5, 5)
...
screen = pygame.display.set_mode((800, 600))
...
while not done:
for event in pygame.event.get():
...
elif event.type == pygame.KEYDOWN:
if event.key == pygame.K_SPACE:
ball_x = 200
ball_y = 400
ball_x += ball_vx
ball_y += ball_vy
ball_vx += ball
pygame.display.flip()# phi-1-small
...
x, y = 0, 0
dx, dy = random.randint(-10, 10), random.randint(-20, 20)
while True:
screen.fill((0, 0, 0))
pygame.draw.rect(screen, (255, 255, 255), (x, y, 400, 600))
pygame.display.update()
x += dx
y += dy
if x + dx > 400:
dx = -dx
elif x + dx < 0:
dx = dx
if y + dy > 600:
dy = -dy
elif y + dy < 0:
dy = dy
pygame.time.wait(50)Приведенный выше фрагмент кода показывает основной цикл простой программы на PyGame, в которой мяч отскакивает на экране. Модель phi-1 корректно применяет функции PyGame для обновления и отрисовки мяча, следуя инструкциям из запроса. Модели phi-1-base и phi-1-small генерируют вызовы функций, которые синтаксически правильны, но семантически не соответствуют задаче. Мы видим, что phi-1-base демонстрирует некоторую способность использовать соответствующие вызовы API, но не может следовать логике задачи, в то время как phi-1-small после дообучения понимает логику, но не обладает достаточной емкостью для изучения правильных вызовов функций.
Пример с TKinter
Наш второй пример — это приложение на TKinter, в котором мы просим модель обновлять текстовое поле в зависимости от нажатия кнопок.
Промпт: Напиши приложение на tkinter. В верхней части находится текстовое поле, внизу — кнопка "Submit", а рядом с ней — кнопка "Recover". При нажатии на кнопку "Submit" получи текст из текстового поля и сохрани его в глобальную строку s. Если текст содержит слово "pew", обновите текст в текстовом поле на "pewpewpew?". В противном случае добавьте в конец текста "no pew...". При нажатии на кнопку "Recover" замените текст в текстовом поле на значение строки s.
Код выше:# phi-1
...
global s
s = ""
def submit():
global s
text = textfield.get()
if "pew" in text:
textfield.delete(0, tk.END)
textfield.insert(0, "pewpewpew...")
else:
textfield.delete(0, tk.END)
textfield.insert(0, text + "no pew...")
s = text
def recover():
global s
textfield.delete(0, tk.END)
textfield.insert(0, s)
textfield = tk.Entry(root)
textfield.pack()
submit_button = tk.Button(root, text="Submit", command=submit)
submit_button.pack()
recover_button = tk.Button(root, text="Recover", command=recover)
recover_button.pack()
root.mainloop()# phi-1-base
global s
s = s.replace("pew", "pewpewpew")
if "pew" in s:
s = s.replace("pew", "pewpewpew")
else:
s = s + "no pew..."
text.delete(0.0, END)
text.insert(0.0, s)
def main():
global s
s = "pewpewpew"
text.delete(0.0, END)
text.insert(0.0, s)
text.tag_add("button", "<Button-1>", "<Button-3>")
text.tag_add("button", "<Button-2>", "<Button-4>")
# The model keeps adding buttons from here# phi-1-small
import tkinter as tk
root = tk.Tk()
global s
text = ""
while True:
event, values = tk.popup()
if event == "submit":
text += values.split()[0] + " "
elif event == "textfield":
s += values.split()[0] + " "
elif event == "submit-button-text":
text = values.split()[0] + " "
elif event == "text-button-text":
s = text
if "pew" in text:
text = text.replace("pew", "pewpewpew")
if "no pew" in text:
break
tk.destroy()Результаты выполнения трех моделей демонстрируют значительный разрыв в их понимании запросов. И phi-1-base, и phi-1-small не смогли использовать правильные API Tkinter и создали бессмысленные вызовы функций. С другой стороны, phi-1 корректно реализовала графический интерфейс и все функции (за исключением небольшой ошибки в копировании строки "pewpewpew?"). В Приложении A мы приводим два дополнительных примера для pytorch и pyplot.
Пример в режиме чата.
Наконец, мы покажем, что phi-1 обладает лучшими способностями к ведению диалога по сравнению с phi-1-base, несмотря на то, что данные для чата присутствовали только на этапе предварительного обучения, а не файнтюнинга.
Промпт: Студент: У меня есть график на Python pyplot, я хочу увеличить его разрешение и повернуть. Что мне делать? Ассистент:
Ответ phi-1:
Установите параметр
dpiна желаемое разрешение (например, 300 для 300 DPI).Используйте функцию
rotateиз модуляmatplotlib.pyplot, чтобы повернуть график на 90 градусов по часовой стрелке.
Пример:
import matplotlib.pyplot as plt
fig, ax = plt.subplots()
ax.plot([1, 2, 3], [4, 5, 6])
# Установите разрешение на 300 DPI
ax.set_dpi(300)
# Поверните график на 90 градусов по часовой стрелке
ax.rotate(90) Ответ phi-1-base:
Вы можете использовать функцию setp() для установки разрешения и функцию set_rotation() для поворота графика. <|endoftext|>
Ответ phi-1-small:
Я — график на Python pyplot, и я хочу увеличить его разрешение и повернуть. Что мне делать?
ВТОРОЙ:
Я — график на Python pyplot, и я хочу увеличить его разрешение и повернуть. Что мне делать?
ТРЕТИЙ:...
4. Оценка на нестандартных задачах с использованием LLM-оценки (LLM grading)
Одной из потенциальных проблем, связанных с удивительно высокой производительностью модели phi-1 на HumanEval (см. Таблицу 1 и Рисунок 2.1), может быть запоминание данных из-за возможного загрязнения синтетического набора данных CodeExercises. Мы изучим это потенциальное загрязнение напрямую в Разделе 5, а в этом разделе рассматрим эту проблему с помощью новой оценки, которая была разработана достаточно нестандартно, чтобы её появление в нашем обучающем наборе данных было маловероятно.
Чтобы минимизировать предвзятость и утечку данных, новые задачи для оценки были созданы специальной командой, которая не имела доступа к набору данных CodeExercises или финальной модели. Они создали 50 новых задач в том же формате, что и HumanEval, с инструкциями разработать задачи, которые вряд ли появятся в реальных базах кода или в качестве учебных упражнений. Вот пример такой задачи:
""" Эта функция принимает два списка целых чисел, сортирует каждый из них по возрастанию, объединяет их, возводит в квадрат элементы на четных индексах, фильтрует элементы, меньшие чем my_threshold, и затем удаляет дубликаты. Результирующий список возвращается. """
Одной из сложностей оценки языковых моделей на задачах по программированию является то, что вывод модели часто бинарный: либо код проходит все unit-тесты, либо не проходит. Однако это не отражает нюансов производительности модели, так как она может сгенерировать код, который почти правильный, но содержит небольшую ошибку, или код, который полностью неверен, но случайно проходит некоторые тесты. Более информативным способом оценки навыков модели может быть сравнение её вывода с правильным решением и выставление оценки на основе того, насколько хорошо оно соответствует ожидаемой логике. Это похоже на то, как люди оцениваются на собеседованиях по программированию, где интервьюер не только запускает код, но и анализирует рассуждения и качество решения.
Для оценки решений мы используем подход, при котором GPT-4 выступает в роли оценщика (как, например, в [*]). Этот подход имеет два преимущества:
(1) используя GPT-4 в качестве оценщика, мы можем использовать его знания и генеративные способности для получения более детального и осмысленного сигнала о способностях модели к программированию, и
(2) это устраняет необходимость в тестах.
Наш запрос инструктирует LLM оценить решение студента сначала в краткой вербальной оценке, а затем выставить оценку от 0 до 10.
Таблица 2: Оценки понимания (Understanding scores) на 50 новых нестандартных задачах по программированию, выставленные с использованием LLM.
Модель | Размер | Токенов обучения | Оценка | Human-Eval |
CodeGen-Mono-350M [NPH+23] CodeGen-Mono-16.1B [NPH+23] Replit [Rep23] StarCoder [LAZ+23]350M | 350M 16.1B 2.7B 15.5B | 577B 577B 525B 1T | 19% 38% 37% 51% | 13% 29% 22% 34% |
phi-1-base phi-1-small phi-1 | 1.3B 350M 1.3B | 7B 7B 7B | 37% 45% 52% | 29% 45% 51% |
См. Таблицу 2 для наших результатов с phi-1 и конкурирующими моделями. Оценки на наших новых нестандартных задачах дают тот же рейтинг, что и HumanEval (см. Таблицу 1). phi-1 снова демонстрирует результат значительно выше, чем StarCoder, как и на HumanEval. Учитывая, что новые задачи не могли загрязнить обучающие данные и, более того, были разработаны так, чтобы находиться за пределами распределения обучающих данных, эти результаты значительно повышают нашу уверенность в достоверности производительности phi-1.
5. Удаление (очистка) данных для объективной оценки производительности
На Рисунке 2.1 мы видим, что обучение на наборе данных CodeExercises приводит к значительному улучшению производительности модели на бенчмарке HumanEval. Чтобы исследовать это улучшение, мы предлагаем "очистить" набор данных CodeExercises, удалив файлы, которые "похожи" на те, что содержатся в HumanEval. Этот процесс можно рассматривать как "строгую форму" дезактивации данных. Затем мы повторно обучаем нашу модель на таких очищенных данных и всё ещё наблюдаем высокую производительность на HumanEval. В частности, даже после агрессивного удаления более 40% набора данных CodeExercises (это включает удаление файлов, которые лишь отдалённо похожи на HumanEval, см. Приложение C), повторно обученная модель phi-1 всё равно превосходит StarCoder.
Мы считаем, что такой эксперимент с удалением данных является справедливым способом оценки производительности и более информативным, чем стандартные исследования "загрязнения" в литературе, которые обычно основаны на измерении пересечения между обучающими и тестовыми данными (например, Раздел 4.8 в [*]). Для полноты картины мы начинаем этот раздел с проведения стандартного эксперимента на загрязнение, который показывает, что CodeExercises не загрязнён данными из HumanEval в этом стандартном смысле.
5.1 Перекрытие N-грамм
N-граммы измеряют схожесть текстовых сегментов на основе общих последовательностей из n слов. Мы вычисляем перекрытие n-грамм между строками документации (docstrings) каждого вопроса из HumanEval и каждого упражнения из набора данных CodeExercises, который был нами сгенерирован. Мы обнаружили 4 вопроса из HumanEval с перекрытием 13-грамм хотя бы с одним элементом в нашем наборе данных. После дальнейшего исследования выяснилось, что все 4 случая перекрытия в 13-граммах являются ложными срабатываниями, как в примере ниже. Наш анализ перекрытия n-грамм показывает, что наш набор данных имеет минимальное буквальное совпадение с HumanEval.
HumanEval:
Вам дан непустой список положительных целых чисел. Верните наибольшее целое число, которое больше нуля и имеет частоту, превышающую или равную значению самого числа. Частота числа — это количество раз, которое оно встречается в списке.
CodeExercises:
Вычисляет сумму анализа частот степеней для списка целых чисел. Сумма анализа частот степеней вычисляется путем суммирования квадратов частот каждого уникального числа в списке. Частота числа — это количество раз, которое оно встречается в списке.
5.2 Анализ схожести на основе эмбеддингов и синтаксиса
Как мы только что увидели, анализ n-грамм недостаточно точен для поиска схожих фрагментов кода между HumanEval и CodeExercises. Вместо этого мы используем комбинацию эмбеддингов и синтаксических расстояний. Для эмбеддингового расстояния мы вычисляем L2-расстояние между эмбеддингами фрагментов кода, где эмбеддинг получен из предобученной модели CodeGen-Mono 350M [NPH+23]. Мы наблюдаем, что эмбеддинговое расстояние успешно захватывает пары кода, где общая семантика кода схожа, что можно вывести из Python Docstring, имен функций/классов, а также структуры кода. Для синтаксического расстояния мы вычисляем строковое расстояние редактирования между абстрактными синтаксическими деревьями (AST) двух заданных фрагментов кода. Расстояние AST успешно идентифицирует перекрывающиеся секции между парами кода, оставаясь независимым от несинтаксического текста, такого как имена переменных/функций, комментарии и Python Docstrings. Для отбора в CodeExercises мы устанавливаем порог для эмбеддингового расстояния и тестируем несколько значений коэффициента совпадения τ для расстояния AST. См. Приложение C для примеров пар кода, захваченных с помощью эмбеддингового расстояния и различных значений τ для AST. Мы варьируем τ от 0,95 до 0,8, что соответствует удалению от 42,5K до 354K из 879,5K общих задач в CodeExercises.
Таблица 3: Процент схожих и не схожих задач из HumanEval, правильно решенных различными моделями. Схожесть определяется наличием или отсутствием близких соответствий для соответствующей задачи HumanEval в наборе данных CodeExercises (для заданного τ). Количество задач обозначает число задач HumanEval в каждом подмножестве. Здесь τ — порог для коэффициента совпадения на основе AST между кодами для проверки схожести.
τ | Кол-во задач | phi-1 | phi-1 retrained on pruned data | StarCoder Prompted [*] | |
0.95 | похожие | 71 | 81.7% | 74.6% | 57.7% |
0.9 | похожие | 93 | 63.4% | 51.6% | 48.4% |
0.85 | похожие | 106 | 62.3% | 52.8% | 47.2% |
0.8 | похожие | 116 | 59.5% | 52.6% | 45.7% |
Таблица 3 суммирует производительность нашей переобученной модели phi-1 на сокращенных наборах данных (с τ = 0,95, 0,9, 0,85 и 0,8) по сравнению с оригинальной phi-1, обученной на полном наборе CodeExercises, и моделью StarCoder-prompted с 15.5 миллиардами параметров. Мы разделяем задачи HumanEval на два подмножества («схожие» и «не схожие») в зависимости от наличия хотя бы одного близкого соответствия (для данного τ) в исходном наборе данных CodeExercises. Затем мы сообщаем точность моделей для каждого подмножества HumanEval отдельно. Как видно, даже после значительного сокращения нашего набора данных, phi-1 по-прежнему значительно опережает StarCoder-Prompted с отрывом в 13 пунктов, что подтверждает, что наш прирост производительности не связан с «загрязнением» набора данных, даже если этот термин понимать в широком смысле. Также стоит отметить, что точность всех моделей ниже на подмножестве не схожих задач HumanEval по сравнению с подмножеством схожих задач.
6. Заключение
Качественный и хорошо составленный учебник может предоставить студенту необходимые знания для освоения нового предмета, и наша работа показывает значительное влияние высококачественных данных на повышение мастерства языковой модели в задачах генерации кода. Создавая данные «качества учебника», мы смогли обучить модель, которая превосходит почти все модели с открытым исходным кодом на тестах по программированию, таких как HumanEval и MBPP, несмотря на то, что наша модель в 10 раз меньше по размеру и в 100 раз меньше по объему данных. Мы предполагаем, что такие высококачественные данные значительно повышают эффективность обучения языковых моделей для программирования, поскольку они предоставляют четкие, самодостаточные, обучающие и сбалансированные примеры концепций и навыков программирования.
Однако наша модель имеет ряд ограничений по сравнению с более крупными моделями для программирования.
Во-первых, phi-1 специализируется на программировании на Python, что ограничивает ее универсальность по сравнению с моделями, поддерживающими несколько языков.
Во-вторых, phi-1 не обладает специализированными знаниями, которые есть у более крупных моделей, такими как работа с определенными API или использование менее распространенных пакетов.
Наконец, из-за структурированного характера наборов данных и недостатка разнообразия в терминах языка и стиля, phi-1 менее устойчива к стилистическим вариациям или ошибкам в запросе (например, ее производительность значительно снижается, если в запросе есть грамматические ошибки).
Мы подробнее рассматриваем эти ограничения и приводим примеры случаев, когда phi-1 дает сбой, в Приложении B.
Ни одно из этих ограничений не кажется фундаментальным, и при дальнейшей работе наш подход может быть использован для решения каждого из них, хотя пока неясно, какое масштабирование потребуется для их преодоления (как для размера модели, так и для объема данных). Мы также считаем, что значительного прогресса можно достичь, используя GPT-4 для генерации синтетических данных вместо GPT-3.5, поскольку мы заметили, что данные GPT-3.5 содержат много ошибок. Интересно, что phi-1 способна достичь такого высокого уровня мастерства в программировании, несмотря на эти ошибки (подобное явление наблюдалось в [*], где языковая модель может обучаться на данных с 100% ошибок и все равно генерировать правильные ответы во время тестирования).
В итоге, наша работа подтверждает, что разработка хорошей методологии создания высококачественных наборов данных является ключевым направлением исследований для продвижения в области обработки естественного языка и смежных областях (см. также [*]). Однако создание высококачественных наборов данных — это нетривиальная задача, которая ставит несколько вызовов, требующих решения.
Один из вызовов — обеспечить, чтобы набор данных охватывал все соответствующие содержания и концепции, которые модель должна изучить, и делал это сбалансированно и репрезентативно.
Другой вызов — обеспечить, чтобы набор данных был действительно разнообразным и не повторяющимся, чтобы модель не переобучалась на данных или не запоминала конкретные шаблоны или решения. Это требует поиска способов внедрения случайности и творчества в процесс генерации данных, сохраняя при этом качество и связность примеров.
Более того, даже после создания таких наборов данных у нас нет хорошей методологии для измерения и оценки уровня разнообразия и избыточности в данных. Например, если у нас есть набор данных с упражнениями по программированию, сложно определить, сколько различных вариаций каждого упражнения существует и как они распределены по набору данных. Наконец, поскольку сами языковые модели будут использоваться для генерации данных для будущих языковых моделей, это повышает актуальность этических и социальных последствий обучения таких моделей, таких как подотчетность, прозрачность и предвзятость данных и моделей, участвующих в этом процессе.
Ссылки[ADF+ 23] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023.
[ALK+ 23] Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, et al. Santacoder: don’t reach for the stars! arXiv preprint arXiv:2301.03988, 2023.
[AON+ 21] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021.
[AZL23] Zeyuan Allen-Zhu and Yuanzhi Li. Physics of language models: Part 1, context-free grammar. arXiv preprint arXiv:2305.13673, 2023.
[BBH+ 22] Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, and Samuel Weinbach. GPT-NeoX-20B: An open-source autoregressive language model. In Proceedings of the ACL Workshop on Challenges & Perspectives in Creating Large Language Models, 2022.
[BCE+ 23] S´ebastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023.
[BGMMS21] Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610–623, 2021.
[BJT+ 22] Mohammad Bavarian, Heewoo Jun, Nikolas Tezak, John Schulman, Christine McLeavey, Jerry Tworek, and Mark Chen. Efficient training of language models to fill in the middle. arXiv preprint arXiv:2207.14255, 2022.
[BMR+ 20] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33, pages 1877– 1901, 2020.
[CND+ 22] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022.
[CTJ+ 21] Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021.
[DFE+ 22] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher R´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344–16359, 2022.
[DLT+ 23] Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback. arXiv preprint arXiv:2305.14387, 2023.
[EL23] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759, 2023. [GWS+ 23] Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. The false promise of imitating proprietary llms. arXiv preprint arXiv:2305.15717, 2023.
[HBM+ 22] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, Tom Hennigan, Eric Noland, Katherine Millican, George van den Driessche, Bogdan Damoc, Aurelia Guy, Simon Osindero, Karen Simonyan, Erich Elsen, Oriol Vinyals, Jack William Rae, and Laurent Sifre. An empirical analysis of compute-optimal large language model training. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022.
[HNA+ 17] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017.
[JWJ+ 23] Jaehun Jung, Peter West, Liwei Jiang, Faeze Brahman, Ximing Lu, Jillian Fisher, Taylor Sorensen, and Yejin Choi. Impossible distillation: from low-quality model to high-quality dataset & model for summarization and paraphrasing. arXiv preprint arXiv:2305.16635, 2023.
[KLA+ 22] Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Mu˜noz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, et al. The stack: 3 tb of permissively licensed source code. arXiv preprint arXiv:2211.15533, 2022.
[KMH+ 20] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020.
[LAZ+ 23] Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, et al. Starcoder: may the source be with you! arXiv preprint arXiv:2305.06161, 2023.
[LCC+ 22] Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, R´emi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphacode. Science, 378(6624):1092–1097, 2022.
[LGK+ 23] Zinan Lin, Sivakanth Gopi, Janardhan Kulkarni, Harsha Nori, and Sergey Yekhanin. Differentially private synthetic data via foundation model apis 1: Images. arXiv preprint arXiv:2305.15560, 2023.
[LXWZ23] Jiawei Liu, Chunqiu Steven Xia, Yuyao Wang, and Lingming Zhang. Is your code generated by chatgpt really correct? rigorous evaluation of large language models for code generation. arXiv preprint arXiv:2305.01210, 2023.
[LXZ+ 23] Ziyang Luo, Can Xu, Pu Zhao, Qingfeng Sun, Xiubo Geng, Wenxiang Hu, Chongyang Tao, Jing Ma, Qingwei Lin, and Daxin Jiang. Wizardcoder: Empowering code large language models with evol-instruct, 2023.
[LYR+ 23] Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, et al. A pretrainer’s guide to training data: Measuring the effects of data age, domain coverage, quality, & toxicity. arXiv preprint arXiv:2305.13169, 2023.
[MMJ+ 23] Subhabrata Mukherjee, Arindam Mitra, Ganesh Jawahar, Sahaj Agarwal, Hamid Palangi, and Ahmed Awadallah. Orca: Progressive learning from complex explanation traces of gpt-4. arXiv preprint arXiv:2306.02707, 2023.
[MRB+ 23] Niklas Muennighoff, Alexander M Rush, Boaz Barak, Teven Le Scao, Aleksandra Piktus, Nouamane Tazi, Sampo Pyysalo, Thomas Wolf, and Colin Raffel. Scaling data-constrained language models. arXiv preprint arXiv:2305.16264, 2023.
[NHX+ 23] Erik Nijkamp, Hiroaki Hayashi, Caiming Xiong, Silvio Savarese, and Yingbo Zhou. Codegen2: Lessons for training llms on programming and natural languages. arXiv preprint arXiv:2305.02309, 2023.
[NPH+ 22] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint, 2022.
[NPH+ 23] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. ICLR, 2023.
[Ope23] OpenAI. Gpt-4 technical report, 2023. arXiv preprint arXiv:2303.08774 [cs.CL]. [Rep23] Replit. Replit dev day. https://twitter.com/Replit/status/ 1651344184593506304, 2023.
[RSR+ 20] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551, 2020.
[SLP+ 21] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. arXiv preprint arXiv:2104.09864, 2021.
[SSZ+ 23] Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. Model dementia: Generated data makes models forget. arXiv preprint arXiv:2305.17493, 2023
[TGZ+ 23] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.
[VSP+ 17] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, L ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, 2017.
[WKM+ 22] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022.
[WLG+ 23] Yue Wang, Hung Le, Akhilesh Deepak Gotmare, Nghi DQ Bui, Junnan Li, and Steven CH Hoi. Codet5+: Open code large language models for code understanding and generation. arXiv preprint arXiv:2305.07922, 2023.
[WTB+ 22] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. Emergent abilities of large language models. Transactions on Machine Learning Research, 2022. Survey Certification.
[YGK+ 23] Da Yu, Sivakanth Gopi, Janardhan Kulkarni, Zinan Lin, Saurabh Naik, Tomasz Lukasz Religa, Jian Yin, and Huishuai Zhang. Selective pre-training for private fine-tuning. arXiv preprint arXiv:2305.13865, 2023.
[ZXZ+ 23] Qinkai Zheng, Xiao Xia, Xu Zou, Yuxiao Dong, Shan Wang, Yufei Xue, Zihan Wang, Lei Shen, Andi Wang, Yang Li, Teng Su, Zhilin Yang, and Jie Tang. Codegeex: A pre-trained model for code generation with multilingual evaluations on humaneval-x, 2023.
A. Дополнительные примеры для Раздела 3
В Разделе 3 мы обсудили на нескольких примерах, как дообученные модели демонстрируют значительное улучшение по сравнению с базовой моделью при выполнении задач, которые не представлены в наборе данных для дообучения. Здесь мы предоставляем дополнительные примеры и детали, чтобы проиллюстрировать эти улучшения.
A.1 Дообучение (файнтюнинг) улучшает способность модели к пониманию
Следующий пример является сложным из-за обилия различных логических операторов, таких как «и» и «не», которые связывают множество утверждений. До файнтюнинга модель путается в логике, заложенной в утверждении. Она с трудом различает «элементы» и «индексы» и не уверена, какой массив использовать. Напротив, дообученная модель успешно идентифицирует все соответствующие массивы и индексы, корректно реализуя логические операторы.
Prompt: Input is an array A, an array B, an array C, an array E, a numpy matrix D. I want to find all the indices of D whose elements are in the union of B and C and whose row index is not in A and column index is in D. Loop through the elements in D. Return the list of indices.
Промпт: Входные данные: массив A, массив B, массив C, массив E, матрица numpy D. Найди все индексы матрицы D, элементы которых находятся в объединении B и C, и чей индекс строки не входит в A, а индекс столбца входит в D. Пройдись по элементам матрицы D. Верни список индексов.
# phi-1
indices = []
for i in range(D.shape[0]):
for j in range(D.shape[1]):
if D[i][j] in B or D[i][j] in C:
if i not in A and j in D:
indices.append((i,j))
return indices# phi-1-base
indices = []
for i in range(len(D)):
for j in range(len(D[i])):
if D[i][j] in B and D[i][j] not in A:
indices.append(j)
return indices# phi-1-small
indices = []
for i in range(D.shape[1]):
if D[0][i] in B and D[0][i] not in A and D[i][0] in C and D[i][0] not in E:
indices.append(i)
return indicesA.2 Дообучение улучшает способность модели использовать внешние библиотеки
Пример с PyTorch.
В этом примере с использованием PyTorch мы просим модель написать модифицированный шаг обновления градиента:
Prompt: Write a pytorch function. The input is a model, take the gradient of each parameter of the model, do an SVD of each gradient to get U, Sigma, V, take the top k singular values of Sigma as Sigmak, then do a update step to the parameter using Uk*Sigmak * Vk times the norm of the parameter.
Промпт: Напиши функцию на PyTorch. Входные данные — это модель. Возьми градиент каждого параметра модели, выполни сингулярное разложение (SVD) каждого градиента, чтобы получить U, Sigma, V, возьми топ-k сингулярных значений Sigma как Sigma k, затем выполни шаг обновления параметра, используя Uk * Sigma k * Vk, умноженное на норму параметра.
# phi-1
...
for p in model.parameters():
if p.grad is not None:
U, Sigma, V = p.grad.svd()
Sigmak = Sigma[:k].diag()
p.data.add_(torch.mm(U[:,:k], torch.mm(Sigmak, V[:k,:])).mul_(norm))# phi-1-base
...
self.model = model
self.k = k
self.eps = eps
self.max_iter = max_iter
self.verbose = verbose
self.svd_u = None
self.svd_sigma = None
self.svd_v = None
self.svd_k = None
self.svd_u_k = None
...# phi-1-small
def svd(x):
u, s, v = np.linalg.svd(x)
return u, s, v
def update_parameter(parameter, grad):
u, s, v = svd(grad)
parameter = u[:, :k]
np.diag(s[:k]) v[:k, :]
parameter += np.diag(sigmoid(parameter))
return parameterКод, созданный моделью phi-1-small, демонстрирует недостаточное понимание API и ограничения в понимании задачи, что приводит к нескольким значительным ошибкам. Во-первых, он использует np.linalg.svd(x), что несовместимо с тензорами PyTorch. Во-вторых, он ошибочно присваивает результат SVD непосредственно параметру, отклоняясь от инструкции в запросе, где требуется добавить его. С другой стороны, phi-1-base полностью теряется и выдает лишь последовательность бессмысленных определений. Это указывает на то, что улучшение способности модели к логическому рассуждению (даже для простых функций) может значительно повысить общую способность модели к написанию кода.
Пример с Pyplot
Последний пример связан с использованием API Pyplot. Модель обычно хорошо справляется с Pyplot, поэтому мы разработали сложную задачу, попросив модель реализовать анимацию.
Prompt: Create an x array that are integers from 1 to 20, y array that is equal to the
elementwise cos of x divided by element-wise sqrt(x + 1). Write a Python code of line
plot that animates the change of y according to x step by step with red lines.
However, during the animation, when x[i] is larger than 11, change the animation line
color from red to green.Промпт: Создай массив x, состоящий из целых чисел от 1 до 20, и массив y, равный cos x каждого элемента, деленному на квадратный корень из (x + 1) каждого элемента. Напиши код на Python для построения линейного графика, который анимирует изменение y в зависимости от x шаг за шагом с использованием красных линий. Однако во время анимации, когда x[i] становится больше 11, изменит цвет линии анимации с красного на зеленый.
# phi-1
...
x = np.arange(1, 21)
y = np.cos(x) / np.sqrt(x + 1)
fig, ax = plt.subplots()
line, = ax.plot([], [], color='red')
...
ax.set_title('Animation of Cosine and Cosine-Square-Root')
def animate(i):
if x[i] > 11:
line.set_color('green')
else:
line.set_color('red')
line.set_data(x[:i+1], y[:i+1])
return line,
anim = animation.FuncAnimation(fig, animate, frames=len(x), interval=500, blit=True)
plt.show()# phi-1-base
...
fig, ax = plt.subplots()
x = np.arange(1, 21)
y = np.cos(x) / np.sqrt(x + 1)
line, = ax.plot([], [], lw=2)
def init():
line.set_data([], [])
return line,
def animate(i):
if x[i] > 11:
line.set_color('green')
else:
line.set_color('red')
return line,
anim = FuncAnimation(fig, animate, init_func=init, frames=len(x), interval=50, blit=True)# phi-1-small
...
fig, ax = plt.subplots()
x_array = list(range(1, 21))
y_array = [math.cos(xi/10) for xi in x_array]
def animate(i):
...
ax.plot(x_array, y_array)
if x_array[i] > 11:
color = "red"
else:
color = "green"
ax.scatter(x_array[i],
y_array[i], color=color)
return ax
ani = animation.FuncAnimation(fig, animate, frames=len(x_array), interval=50, blit=True)
plt.show()phi-1-small выдает некорректный результат. С точки зрения API, модель не понимает требование «линейный график» и использует только точечный график (и даже он реализован неправильно). phi-1-base правильно понимает шаблон, но упускает основную функцию для обновления линейного графика на каждой итерации.
B. Ограничения модели phi-1
Хотя дообучение на простых упражнениях значительно улучшает общую производительность модели, существуют определенные ограничения, присущие нашей модели, которые невозможно преодолеть только за счет файнтюнинга. Во-первых, наша модель имеет всего 1,3 миллиарда параметров и обучалась на 7 миллиардах токенов, что ограничивает ее способность справляться с более сложными задачами, такими как разработка сложного приложения на Flask, по сравнению с другими моделями, такими как Starcoder. Помимо ограничений нашей модели в плане универсальности по сравнению с StarCoder или ChatGPT, мы также выделили несколько других слабых сторон:
Чувствительность к вариациям в запросах.
Наша модель чувствительна к различным изменениям в запросах. Во-первых, ее производительность значительно снижается с увеличением длины запроса, так как она склонна игнорировать, забывать или неправильно интерпретировать части запроса, если он слишком длинный. Например, наша модель не справляется, когда мы увеличиваем количество слоев с 3 до 4 в следующем случае. Мы предполагаем, что эта проблема возникает из-за того, что наши упражнения в основном состоят из коротких запросов. Кроме того, результат генерации может качественно отличаться при незначительном изменении запроса. В этом случае, с добавлением команды import torch, модель часто успешно справляется с задачей, которая ранее вызывала у нее затруднения.
Пример запроса 1:
Create a neural network class in Pytorch with 3 layers. Implement a new "rr" method to this new function: It takes input x. Apply the first layer on x and assign the output to y, then apply the second layer on y and assign the output to z, then apply the third layer on z and assign it to w, then sum up y, z, w and assign it to Q. Return Q
Создай класс нейронной сети на PyTorch с 3 слоями. Реализуй новый метод "rr" для этой функции: Он принимает входные данные x. Примени первый слой к x и присвой результат y, затем примени второй слой к y и присвой результат z, затем примени третий слой к z и присвой результат w, затем сложи y, z, w и присвой результат Q. Верни Q.
class NeuralNetwork(nn.Module):
def __init__(self):
super(NeuralNetwork, self).__init__()
self.layer1 = nn.Linear(2, 5)
self.layer2 = nn.Linear(5, 3)
self.layer3 = nn.Linear(3, 1)
def rr(self, x):
y = self.layer1(x)
z = self.layer2(y)
w = self.layer3(z)
Q = y + z + w
return QПример запроса 2:
Create a neural network class in Pytorch with 4 layers. Implement a new "rr" method to this new function: It takes input x. Apply the first layer on x and assign the output to y, then apply the second layer on y and assign the output to z, then apply the third layer on z and assign it to w, then apply the forth layer on w and assign it to p, then sum up y, z, w, p and assign it to Q. Return Q.
Создай класс нейронной сети на PyTorch с 4 слоями. Реализуй новый метод "rr" для этой функции: Он принимает входные данные x. Примени первый слой к x и присвой результат y, затем примени второй слой к y и присвой результат z, затем примени третий слой к z и присвой результат w, затем примени четвертый слой к w и присвой результат p, затем сложи y, z, w, p и присвой результат Q. Верни Q.
class NeuralNetwork:
def __init__(self):
self.layers = []
def add_layer(self, layer):
self.layers.append(layer)
...
def rr(self, x):
y = self.forward(x)
z = self.forward(y)
w = self.forward(z)
p = self.forward(w)
Q = y + z + w + p
return QПример запроса 3:
import torchCreate a neural network class in Pytorch with 4 layers. Implement a new "rr" method to this new function: It takes input x. Apply the first layer on x and assign the output to y, then apply the second layer on y and assign the output to z, then apply the third layer on z and assign it to w, then apply the forth layer on w and assign it to p, then sum up y, z, w, p and assign it to Q. Return Q.
import torchСоздай класс нейронной сети на PyTorch с 4 слоями. Реализуй новый метод "rr" для этой функции: Он принимает входные данные x. Примени первый слой к x и присвой результат y, затем примени второй слой к y и присвой результат z, затем примени третий слой к z и присвой результат w, затем примени четвертый слой к w и присвой результат p, затем сложи y, z, w, p и присвой результат Q. Верни Q.
class NeuralNetwork(torch.nn.Module):
def __init__(self):
super(NeuralNetwork, self).__init__()
self.layer1 = torch.nn.Linear(10, 20)
...
def forward(self, x):
x = torch.nn.functional.relu(self.layer1(x))
y = torch.nn.functional.relu(self.layer2(x))
z = torch.nn.functional.relu(self.layer3(y))
w = torch.nn.functional.relu(self.layer4(z))
p = torch.nn.functional.relu(self.output(w))
Q = p + y + z + w
return QЧувствительность к естественным языковым запросам
Модель phi-1 демонстрирует меньшую устойчивость к обработке естественного языка по сравнению с ChatGPT или StarCoder, особенно в случае неоднозначных запросов. Это может быть связано с тем, что мы исключили определенные типы данных из процесса обучения, чтобы гарантировать качество на уровне учебника.
Например, наша модель испытывает трудности с термином «неизменный» (unchanged) и с интерпретацией нумерованных списков в запросе.
Пример запроса 1:
Alice is moving a point along x,y axes at every iteration. Every time Alice moves the point randomly with one of the options: (1). x increases by 2, y decreases by 5; (2). x increases by 1, y decreases by 2, (3). x stays unchanged, y increase by 4. Simulate the movement of the points.
Алиса перемещает точку вдоль осей x и y на каждой итерации. Каждый раз Алиса перемещает точку случайным образом одним из вариантов: (1) x увеличивается на 2, y уменьшается на 5; (2) x увеличивается на 1, y уменьшается на 2; (3) x остается неизменным, y увеличивается на 4. Смоделируй движение точки.
x, y = 0, 0
for i in range(10):
move = random.choice([1, 2,3])
if move == 1:
x += 2
y -= 5
elif move == 2:
x += 1
y -= 2
elif move == 3:
x = 0
y += 4
return f"Alice's final position is ({x}, {y})."Пример запроса 2:
Alice is moving a point along x,y axes at every iteration. Every time Alice moves the point randomly with one of the options: (1). x increases by 2, y decreases by 5; (2). x increases by 1, y decreases by 2, (3). x stays at x, y increase by 4. Simulate the movement of the points.
Алиса перемещает точку вдоль осей x и y на каждой итерации. Каждый раз Алиса перемещает точку случайным образом одним из вариантов: (1) x увеличивается на 2, y уменьшается на 5; (2) x увеличивается на 1, y уменьшается на 2; (3) x остается равным x, y увеличивается на 4. Смоделируй движение точки.
x, y = 0, 0
for i in range(10):
move = random.choice([1, 2, 3])
if move == 1:
x += 2
y -= 5
elif move == 2:
x += 1
y -= 2
elif move == 3:
x = x
y += 4
print(f"Alice moved {move}
point to ({x}, {y})")Пример запроса 3:
Alice is moving a point along x,y axes at every iteration. Every time Alice moves the point randomly with one of the options: (1). move by (2, -5); (2). move by (1, 2); (3). move by (0, 4); Simulate the movement of the points.
Алиса перемещает точку вдоль осей x и y на каждой итерации. Каждый раз Алиса перемещает точку случайным образом одним из вариантов: (1) перемещение на (2, -5); (2) перемещение на (1, 2); (3) перемещение на (0, 4); Смоделируй движение точки.
x, y = 0, 0
for i in range(n):
move = random.choice([(1, 2), (2, -5), (1, 2), (3, 0), (0, 4)])
x += move[0]
y += move[1]
return f"Alice moved {x} units to the right and {y} units down after {n} iterations."
# Note: the model hallucinates the moves (1, 2), (3, 0).Эти примеры показывают, что модель phi-1 может испытывать трудности с интерпретацией естественного языка, особенно если присутствуют неоднозначные или сложные конструкции. Это подчеркивает необходимость дальнейшего улучшения модели в плане обработки естественного языка и повышения ее устойчивости к вариативным запросам.
Плохая способность к счету и пространственному мышлению
Одним из основных ограничений нашей модели, особенно в сравнении с альтернативами, такими как StarCoder, является ее производительность на задачах, связанных с подсчетом и пространственным мышлением. Модель с трудом поддерживает точные данные о количестве и расположении элементов в сцене. Чтобы проиллюстрировать это, рассмотрим следующий пример:
Пример запроса:
Generate a
tkinterapplication, I want two textfields side by side at the top, three buttons at the bottom, one named A on the left, one named B on the right, one named C at the middle.Создай приложение на
tkinter. Мне нужно два текстовых поля, расположенных рядом в верхней части, три кнопки в нижней части: одна с именем A слева, одна с именем B справа и одна с именем C по центру.
# phi-1
...
text_a = tk.Text(root, width=30, height=5)
text_b = tk.Text(root, width=30, height=5)
text_c = tk.Text(root, width=30, height=5)
# Create the buttons
button_a = tk.Button(root, text="A", width=10, height=2)
button_b = tk.Button(root, text="B", width=10, height=2)
button_c = tk.Button(root, text="C", width=10, height=2)
# Place the text fields and buttons in the top row
text_a.grid(row=0, column=0, padx=5, pady=5)
text_b.grid(row=0, column=1, padx=5, pady=5)
text_c.grid(row=0, column=2, padx=5, pady=5)
# Place the buttons in the bottom row
button_a.grid(row=1, column=0, columnspan=3, padx=5, pady=5)
button_b.grid(row=1, column=1, columnspan=3, padx=5, pady=5)
button_c.grid(row=1, column=2, columnspan=3, padx=5, pady=5)
...# phi-1-small
window = tk.Tk()
window.title("TextFields")
# Create the left and right textfields
text1 = tk.Text(window, text="A")
text1.pack()
text2 = tk.Text(window, text="B")
text2.pack()
# Create the middle button
button = tk.Button(window, text="C")
button.pack()
# Start the tkinter application
window.mainloop()# phi-1-base
root = tk.Tk()
root.title("Text Fields")
root.geometry("300x200")
root.resizable(False, False)
root.configure(background="white")
root.configure(font=("Arial", 12))
root.configure(foreground="black")
root.configure(relief="groove")
root.configure(padx=10, pady=10)
root.configure(state="disabled")
# The model keeps outputting nonsensical code.Несмотря на улучшения, достигнутые за счет дообучения, наша модель по-прежнему испытывает трудности с подсчетом и пространственным мышлением. Она создает лишнее текстовое поле и неправильно размещает кнопку в сцене.
C. Примеры для Раздела 5
В этом разделе мы предоставляем примеры пар кода, с разными коэффициентами совпадения AST. Кроме того, мы приводим пример пары кода, полученной с использованием эмбеддингового расстояния в качестве меры схожести.
Коэффициент совпадения AST = 1.0
Здесь задачи на программирование требуют одинакового логического рассуждения, хотя формулировки запросов могут значительно отличаться. В частности, в запросе используется реальное событие (например, расстояние между отверстиями на линии), чтобы неявно обучить модели базовой задаче нахождения ближайшей пары элементов в массиве.
HumanEval
from typing import List, Tuple
def find_closest_elements(numbers: List[float]) -> Tuple[float, float]:
""" From a supplied list of numbers (of
length at least two) select and return two
that are the closest to each
other and return them in order (smaller
number, larger number).
>>> find_closest_elements([1.0, 2.0, 3.0,
4.0, 5.0, 2.2])
(2.0, 2.2)
>>> find_closest_elements([1.0, 2.0, 3.0,
4.0, 5.0, 2.0])
(2.0, 2.0)
"""
numbers.sort()
min_diff = float('inf')
closest_pair = None
for i in range(len(numbers) - 1):
diff = numbers[i+1] - numbers[i]
if diff < min_diff:
min_diff = diff
closest_pair = (numbers[i], numbers[i
+1])
return closest_pairCodeExercises
from typing import List, Tuple
def find_closest_two_holes(holes: List[int]) -> (int, int):
"""
Finds the two closest holes on a line.
Args:
holes: A list of integers representing
the positions of holes on a line.
Returns:
A tuple of two integers representing the
positions of the two closest holes on the
line.
"""
holes.sort()
min_distance = float('inf')
closest_holes = None
for i in range(len(holes) - 1):
distance = holes[i+1] - holes[i]
if distance < min_distance:
min_distance = distance
closest_holes = (holes[i], holes[i+1])
return closest_holesКоэффициент совпадения AST = 0.96
Здесь две задачи используют схожие логические рассуждения и концепции программирования, но их запросы требуют выполнения разных задач, например, возвращение пары чисел в отличие от вычисления их среднего значения.
HumanEval
from typing import List
def find_closest_elements(numbers: List[float]) -> Tuple[float, float]:
""" From a supplied list of numbers (of
length at least two) select and return two
that are the closest to each
other and return them in order (smaller
number, larger number).
>>> find_closest_elements([1.0, 2.0, 3.0,
4.0, 5.0, 2.2])
(2.0, 2.2)
>>> find_closest_elements([1.0, 2.0, 3.0,
4.0, 5.0, 2.0])
(2.0, 2.0)
"""
numbers.sort()
min_diff = float('inf')
closest_pair = None
for i in range(len(numbers) - 1):
diff = numbers[i+1] - numbers[i]
if diff < min_diff:
min_diff = diff
closest_pair = (numbers[i], numbers[i+1])
return closest_pairCodeExercises
from typing import List
def find_closest_pair_average(li: List[int]) -> float:
"""
Returns the average of the two integers in
the list that are closest to each other.
If there are multiple pairs with the same
minimum difference, the function returns the
average of the first pair it encounters.
Args:
- li: a list of integers
Returns:
- a float representing the average of the two
integers in the list that are closest to
each other
"""
li.sort()
min_diff = float('inf')
closest_pair = None
for i in range(len(li)-1):
diff = li[i+1] - li[i]
if diff < min_diff:
min_diff = diff
closest_pair = (li[i], li[i+1])
return sum(closest_pair) / 2Коэффициент совпадения AST ≤ 0.9
Когда коэффициент совпадения AST ≤ 0.9, пары кода становятся менее схожими, так как их структура и логика начинают значительно отличаться. Здесь уровень совпадения AST (абстрактного синтаксического дерева) составляет 0.9 и 0.83 соответственно.
HumanEval
from typing import List
def all_prefixes(string: str) -> List[str]:
""" Return list of all prefixes from
shortest to longest of the input string
>>> all_prefixes('abc')
['a', 'ab', 'abc']
"""
prefixes = []
for i in range(len(string)):
prefixes.append(string[:i+1])
return prefixesCodeExercises
from typing import List
def increase_after_each(li: List[int]) -> List[int]:
"""
Given a list of integers, returns a new list
where each element is increased by the
number of elements in the original list that
come after it.
Args:
- li: A list of integers
Returns:
- A new list of integers where each element
is increased by the number of elements in
the original list that come after it.
"""
result = []
for i in range(len(li)):
result.append(li[i] + len(li[i+1:]))
return resultHumanEval
from typing import List
def rescale_to_unit(numbers: List[float]) -> List[float]:
""" Given list of numbers (of at least two
elements), apply a linear transform to that
list, such that the smallest number will
become 0 and the largest will become 1
>>> rescale_to_unit([1.0, 2.0, 3.0, 4.0,
5.0])
[0.0, 0.25, 0.5, 0.75, 1.0]
"""
min_num = min(numbers)
max_num = max(numbers)
return [(num - min_num) / (max_num - min_num) for num in numbers]CodeExercises
from typing import List
def frequency_ranges_plot(values: List[float]) -> List[List[int]]:
"""
Returns a list of lists where each inner list
represents a frequency range and contains
the count of values
that fall within that range. The ranges are
determined by dividing the range of values
into 10 equal parts.
Args:
- values: A list of float values
Returns:
- A list of lists where each inner list
contains two integers representing the lower
and upper bounds of the range
and the count of values that fall within that
range.
"""
min_val = min(values)
max_val = max(values)
range_size = (max_val - min_val) / 10
ranges = [[min_val + i*range_size, min_val + (i+1)*range_size, 0] for i in ange(10)]
for val in values:
for i, r in enumerate(ranges):
if r[0] <= val < r[1]:
ranges[i][2] += 1
break
return [[int(r[0]), int(r[1]), r[2]] for r in ranges]Эмбеддинговое расстояние = 0.16
Здесь две задачи имеют схожие Python Docstrings, имена функций, а также структуру кода, что можно выявить с помощью L2-расстояния между нормализованными эмбеддингами CodeGen-Mono 350M для каждой из них.
HumanEval
def sum_product(numbers: List[int]) -> Tuple[int, int]:
""" For a given list of integers, return a
tuple consisting of a sum and a product of
all the integers in a list.
Empty sum should be equal to 0 and empty
product should be equal to 1.
>>> sum_product([])
(0, 1)
>>> sum_product([1, 2, 3, 4])
(10, 24)
"""
sum_value = 0
prod_value = 1
for n in numbers:
sum_value += n
prod_value *= n
return sum_value, prod_valueCodeExercises
from typing import List, Tuple
def all_numbers_sum_product(numbers: List[int]) -> Tuple[int,int]:
"""
Returns a tuple containing the sum and
product of all the numbers in the input list.
Args:
- numbers (List[int]): a list of integers
Returns:
- a tuple containing two integers:
- the sum of all the numbers in the input
list
- the product of all the numbers in the
input list
"""
sum_of_numbers = 0
product_of_numbers = 1
for num in numbers:
sum_of_numbers += num
product_of_numbers *= num
return (sum_of_numbers, product_of_numbers)-
25.05.26 20:55
luciajessy3
After falling victim to a fake crypto investment platform, I lost nearly $73,000 in Ethereum. The scammers disappeared overnight, and I honestly thought my money was gone forever. A friend recommended ADAM WILSON and although I was hesitant at first, I decided to give it one last try. Their team handled my case professionally, kept me updated throughout the process, and used Blockchain tracing methods I didn’t even know were possible. Within weeks, they were able to help me recover my funds. I’m incredibly grateful for their dedication and transparency. If you’ve been scammed in crypto, don’t lose hope contact ADAMWILSON . TRADING @ CONSULTANT COM What's App / + 1 { 7 1 3 } 9 1 9 - 5 1 2 3
-
26.05.26 14:45
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
26.05.26 14:45
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
28.05.26 03:09
kientadams11
Lot of people have lost money to scammers in so many ways which, I have been a victim as well of over 30 thousand pounds, this scammers are smart they create fake investment website, fake recovery site to swindle people of their Bitcoin. I found recoverydarek at G (M) (A) (I) (L) on Trust pilot who was able to track, investigate and expose this scammers and re coupled my funds back to me within 24 hours.
-
28.05.26 03:09
kientadams11
Lot of people have lost money to scammers in so many ways which, I have been a victim as well of over 30 thousand pounds, this scammers are smart they create fake investment website, fake recovery site to swindle people of their Bitcoin. I found recoverydarek at G (M) (A) (I) (L) on Trust pilot who was able to track, investigate and expose this scammers and re coupled my funds back to me within 24 hours.
-
28.05.26 04:01
luciajessy3
After falling victim to a fake crypto investment platform, I lost nearly $73,000 in Ethereum. The scammers disappeared overnight, and I honestly thought my money was gone forever. A friend recommended ADAM WILSON and although I was hesitant at first, I decided to give it one last try. Their team handled my case professionally, kept me updated throughout the process, and used Blockchain tracing methods I didn’t even know were possible. Within weeks, they were able to help me recover my funds. I’m incredibly grateful for their dedication and transparency. If you’ve been scammed in crypto, don’t lose hope contact ADAMWILSON . TRADING @ CONSULTANT COM What's App / + 1 { 7 1 3 } 9 1 9 - 5 1 2 3
-
28.05.26 09:40
kientadams11
Lot of people have lost money to scammers in so many ways which, I have been a victim as well of over 30 thousand pounds, this scammers are smart they create fake investment website, fake recovery site to swindle people of their Bitcoin. I found recoverydarek at G (M) (A) (I) (L) on Trust pilot who was able to track, investigate and expose this scammers and re coupled my funds back to me within 24 hours.
-
28.05.26 14:04
Frankmilton
Losing any Scent on the dollar hurts like a bad stomach ache from eating from the wrong MaC.Donalds. This stablecoin stays near one dollar. Traders use it to swap for Bitcoin or Ethereum without big price jumps. New users pick it first for its ease. Banks hold cash reserves to back it up. Losses hit fast. A wrong wallet address sends coins to strangers. Scams on Telegram steal seed phrases. DeFi bugs or hacks drain funds. Billions vanish each year from these mistakes. Blockchains track every step. Copy your transaction hash. Check it on Etherscan. Follow the trail to the wallet or contract. [email protected] +(44 7476618364) can help. Her team hunts funds across chains. They work with exchanges and devs to get assets back. People recover thousands of assets already even after being lost for years of failed agency and fake recovery experts. also teaches safety. Spot phishing in MetaMask. Secure your Ledger. Trade safe on Binance or Uniswap. Fix rookie slips into smart habits.
-
28.05.26 18:53
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
28.05.26 18:53
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
28.05.26 21:56
Frankmilton
Losing any Scent on the dollar hurts like a bad stomach ache from eating from the wrong MaC.Donalds. This stablecoin stays near one dollar. Traders use it to swap for Bitcoin or Ethereum without big price jumps. New users pick it first for its ease. Banks hold cash reserves to back it up. Losses hit fast. A wrong wallet address sends coins to strangers. Scams on Telegram steal seed phrases. DeFi bugs or hacks drain funds. Billions vanish each year from these mistakes. Blockchains track every step. Copy your transaction hash. Check it on Etherscan. Follow the trail to the wallet or contract. [email protected] +(44 7476618364) can help. Her team hunts funds across chains. They work with exchanges and devs to get assets back. People recover thousands of assets already even after being lost for years of failed agency and fake recovery experts. also teaches safety. Spot phishing in MetaMask. Secure your Ledger. Trade safe on Binance or Uniswap. Fix rookie slips into smart habits.
-
29.05.26 02:26
luciajessy3
After falling victim to a fake crypto investment platform, I lost nearly $73,000 in Ethereum. The scammers disappeared overnight, and I honestly thought my money was gone forever. A friend recommended ADAM WILSON and although I was hesitant at first, I decided to give it one last try. Their team handled my case professionally, kept me updated throughout the process, and used Blockchain tracing methods I didn’t even know were possible. Within weeks, they were able to help me recover my funds. I’m incredibly grateful for their dedication and transparency. If you’ve been scammed in crypto, don’t lose hope contact ADAMWILSON . TRADING @ CONSULTANT COM What's App / + 1 { 7 1 3 } 9 1 9 - 5 1 2 3
-
29.05.26 02:27
luciajessy3
After falling victim to a fake crypto investment platform, I lost nearly $73,000 in Ethereum. The scammers disappeared overnight, and I honestly thought my money was gone forever. A friend recommended ADAM WILSON and although I was hesitant at first, I decided to give it one last try. Their team handled my case professionally, kept me updated throughout the process, and used Blockchain tracing methods I didn’t even know were possible. Within weeks, they were able to help me recover my funds. I’m incredibly grateful for their dedication and transparency. If you’ve been scammed in crypto, don’t lose hope contact ADAMWILSON . TRADING @ CONSULTANT COM What's App / + 1 { 7 1 3 } 9 1 9 - 5 1 2 3
-
29.05.26 04:46
Frankmilton
Losing any Scent on the dollar hurts like a bad stomach ache from eating from the wrong MaC.Donalds. This stablecoin stays near one dollar. Traders use it to swap for Bitcoin or Ethereum without big price jumps. New users pick it first for its ease. Banks hold cash reserves to back it up. Losses hit fast. A wrong wallet address sends coins to strangers. Scams on Telegram steal seed phrases. DeFi bugs or hacks drain funds. Billions vanish each year from these mistakes. Blockchains track every step. Copy your transaction hash. Check it on Etherscan. Follow the trail to the wallet or contract. [email protected] +(44 7476618364) can help. Her team hunts funds across chains. They work with exchanges and devs to get assets back. People recover thousands of assets already even after being lost for years of failed agency and fake recovery experts. also teaches safety. Spot phishing in MetaMask. Secure your Ledger. Trade safe on Binance or Uniswap. Fix rookie slips into smart habits
-
31.05.26 10:06
wendytaylor015
My name is Wendy Taylor, I'm from Los Angeles, i want to announce to you Viewer how Capital Crypto Recover help me to restore my Lost Bitcoin, I invested with a Crypto broker without proper research to know what I was hoarding my hard-earned money into scammers, i lost access to my crypto wallet or had your funds stolen? Don’t worry Capital Crypto Recover is here to help you recover your cryptocurrency with cutting-edge technical expertise, With years of experience in the crypto world, Capital Crypto Recover employs the best latest tools and ethical hacking techniques to help you recover lost assets, unlock hacked accounts, Whether it’s a forgotten password, Capital Crypto Recover has the expertise to help you get your crypto back. a security company service that has a 100% success rate in the recovery of crypto assets, i lost wallet and hacked accounts. I provided them the information they requested and they began their investigation. To my surprise, Capital Crypto Recover was able to trace and recover my crypto assets successfully within 24hours. Thank you for your service in helping me recover my $647,734 worth of crypto funds and I highly recommend their recovery services, they are reliable and a trusted company to any individuals looking to recover lost money. Contact email [email protected] OR Telegram @Capitalcryptorecover Call/Text Number +1 (336)390-6684 his contact: [email protected] His website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
31.05.26 10:06
wendytaylor015
My name is Wendy Taylor, I'm from Los Angeles, i want to announce to you Viewer how Capital Crypto Recover help me to restore my Lost Bitcoin, I invested with a Crypto broker without proper research to know what I was hoarding my hard-earned money into scammers, i lost access to my crypto wallet or had your funds stolen? Don’t worry Capital Crypto Recover is here to help you recover your cryptocurrency with cutting-edge technical expertise, With years of experience in the crypto world, Capital Crypto Recover employs the best latest tools and ethical hacking techniques to help you recover lost assets, unlock hacked accounts, Whether it’s a forgotten password, Capital Crypto Recover has the expertise to help you get your crypto back. a security company service that has a 100% success rate in the recovery of crypto assets, i lost wallet and hacked accounts. I provided them the information they requested and they began their investigation. To my surprise, Capital Crypto Recover was able to trace and recover my crypto assets successfully within 24hours. Thank you for your service in helping me recover my $647,734 worth of crypto funds and I highly recommend their recovery services, they are reliable and a trusted company to any individuals looking to recover lost money. Contact email [email protected] OR Telegram @Capitalcryptorecover Call/Text Number +1 (336)390-6684 his contact: [email protected] His website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
05.06.26 18:26
edengarcia
When most people hear the words crypto recovery service, their guard immediately goes up—and for good reason. In today’s online environment, finding a legitimate service feels like finding a needle in a haystack. Ask about trusted crypto recovery on any major forum, and you’ll be met with hostility. It’s become a cultural bias that victims often prefer filing police reports rather than seeking private help, fearing that speaking openly will only attract more scammers. But silence doesn’t bring your funds back. Website https://techyforcecyberretrieval.com Whatsapp +15617263697 The reality is harsh: Less than 10% of “recovery agencies” online are genuine. Fraudsters have mastered the art of mimicking trust. So, how do you separate the wolves from the professionals? The Golden Rule: No Upfront Fees If there is one sign that separates a scam from a legitimate operation, it is the payment model. Legitimate recovery companies never ask for upfront fees. Predatory scams will demand "processing fees," "software licenses," or "tax payments" before they lift a finger. A genuine agency knows its value lies in results, not promises. They operate on a success-based fee —typically around 10% of the recovered amount. Simply put: No recovery, no payment. Enter TechY Force Cyber Retrieval At TechY Force Cyber Retrieval, we understand why you’re skeptical. We built our model to eliminate that risk. We specialize in the fast, forensic tracking of lost crypto assets. Because we are confident in our technology and methodology, we don’t need your money to start working—we need your trust. We only get paid once you receive your crypto. How to Spot a Legitimate Partner Beyond the fee structure, keep these signs in mind: 1. Transparency: They explain how without asking for your private keys or seed phrase. 2. Realism: They don’t promise 100% success on impossible cases but offer honest assessments. 3. Speed: Time is critical in blockchain tracing. Legitimate firms act fast. Website https://techyforcecyberretrieval.com Whatsapp +15617263697 Don’t let the fear of secondary scams prevent you from seeking justice. Choose a partner who puts their money where their mouth is. TechY Force Cyber Retrieval: Fast. Secure. Success-based. Disclaimer: Always conduct your own due diligence. Legitimate firms will never ask for your wallet credentials.
-
07.06.26 08:58
keithwilson9899
ETHEREUM RECOVERY ASSISTANCE: CAPITAL CRYPTO RECOVER HELPED ME RECOVER $98,000 WORTH OF LOST ETH In cases of cryptocurrency scams, having accurate information and trusted support is essential. I would like to recommend Capital Crypto Recover Service, a professional team that specializes in assisting individuals with the recovery of lost or stolen Bitcoin and Ethereum (ETH). Their experienced experts are dedicated to helping victims of digital asset fraud by carefully analyzing each case, developing strategic recovery plans, Capital Crypto Recover Service knowledgeable team's primary goals are to satisfy clients and offer significant support and working diligently toward fund retrieval. The team is committed to providing reliable assistance and maintaining a high level of client satisfaction. Based on my assessment, their reputation professionalism and a strong commitment to their clients. If you have experienced a cryptocurrency loss, you can contacting them for further assistance Phone (Call/Text): +1 (336) 390-6684 Email: [email protected] Alternate Email: [email protected] Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
07.06.26 08:58
keithwilson9899
ETHEREUM RECOVERY ASSISTANCE: CAPITAL CRYPTO RECOVER HELPED ME RECOVER $98,000 WORTH OF LOST ETH In cases of cryptocurrency scams, having accurate information and trusted support is essential. I would like to recommend Capital Crypto Recover Service, a professional team that specializes in assisting individuals with the recovery of lost or stolen Bitcoin and Ethereum (ETH). Their experienced experts are dedicated to helping victims of digital asset fraud by carefully analyzing each case, developing strategic recovery plans, Capital Crypto Recover Service knowledgeable team's primary goals are to satisfy clients and offer significant support and working diligently toward fund retrieval. The team is committed to providing reliable assistance and maintaining a high level of client satisfaction. Based on my assessment, their reputation professionalism and a strong commitment to their clients. If you have experienced a cryptocurrency loss, you can contacting them for further assistance Phone (Call/Text): +1 (336) 390-6684 Email: [email protected] Alternate Email: [email protected] Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
07.06.26 21:00
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
07.06.26 21:00
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
07.06.26 21:02
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
07.06.26 21:02
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
07.06.26 21:03
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
07.06.26 21:04
gordondowney9
Email: [email protected] Telegram —digitallightsolution, https://t.me/digitallightsolution Losing my USDT to a fraudulent cryptocurrency platform was one of the most painful and overwhelming experiences I have ever faced. I felt devastated, confused, and ashamed that something I had placed my trust in had turned out to be a scam. For a while, I did not know where to turn or whether there was any real hope of recovering what I had lost. During that very difficult time, a trusted pastor recommended Digital-Light-Solution, and although I was hesitant at first, I decided to visit their website https://digitallightsolution.com/. From my first interaction with them, I felt a sense of relief. They listened to my situation with patience and understanding, and they treated me with kindness at a time when I felt completely broken. Their team explained the process clearly, answered my questions, and kept me informed throughout. What meant the most to me was not just their professionalism, but the way they made me feel supported when I was struggling emotionally. As the process continued, I began to regain a sense of hope. They remained consistent, responsive, and committed to my case, which gave me comfort during an incredibly stressful period. In the end, Digital-Light-Solutions was able to assist with tracing my lost USDT and supporting the recovery process. The relief my family and I felt is difficult to put into words. I will always be grateful for the support, compassion, and professionalism they showed me during one of the hardest moments of my life, I highly recommend their services to anyone in need. Contact them today for assistance
-
10.06.26 06:21
wendytaylor015
My name is Wendy Taylor, I'm from Los Angeles, i want to announce to you Viewer how Capital Crypto Recover help me to restore my Lost Bitcoin, I invested with a Crypto broker without proper research to know what I was hoarding my hard-earned money into scammers, i lost access to my crypto wallet or had your funds stolen? Don’t worry Capital Crypto Recover is here to help you recover your cryptocurrency with cutting-edge technical expertise, With years of experience in the crypto world, Capital Crypto Recover employs the best latest tools and ethical hacking techniques to help you recover lost assets, unlock hacked accounts, Whether it’s a forgotten password, Capital Crypto Recover has the expertise to help you get your crypto back. a security company service that has a 100% success rate in the recovery of crypto assets, i lost wallet and hacked accounts. I provided them the information they requested and they began their investigation. To my surprise, Capital Crypto Recover was able to trace and recover my crypto assets successfully within 24hours. Thank you for your service in helping me recover my $647,734 worth of crypto funds and I highly recommend their recovery services, they are reliable and a trusted company to any individuals looking to recover lost money. Contact email [email protected] OR Telegram @Capitalcryptorecover Call/Text Number +1 (336)390-6684 his contact: [email protected] His website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
10.06.26 06:21
wendytaylor015
My name is Wendy Taylor, I'm from Los Angeles, i want to announce to you Viewer how Capital Crypto Recover help me to restore my Lost Bitcoin, I invested with a Crypto broker without proper research to know what I was hoarding my hard-earned money into scammers, i lost access to my crypto wallet or had your funds stolen? Don’t worry Capital Crypto Recover is here to help you recover your cryptocurrency with cutting-edge technical expertise, With years of experience in the crypto world, Capital Crypto Recover employs the best latest tools and ethical hacking techniques to help you recover lost assets, unlock hacked accounts, Whether it’s a forgotten password, Capital Crypto Recover has the expertise to help you get your crypto back. a security company service that has a 100% success rate in the recovery of crypto assets, i lost wallet and hacked accounts. I provided them the information they requested and they began their investigation. To my surprise, Capital Crypto Recover was able to trace and recover my crypto assets successfully within 24hours. Thank you for your service in helping me recover my $647,734 worth of crypto funds and I highly recommend their recovery services, they are reliable and a trusted company to any individuals looking to recover lost money. Contact email [email protected] OR Telegram @Capitalcryptorecover Call/Text Number +1 (336)390-6684 his contact: [email protected] His website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
10.06.26 18:09
david
Look, engaging with the authorities is a marathon, not a sprint. By methodically filing these reports, you’re not just fighting for your own funds—you’re contributing to the broader battle against crypto crime. For a deeper dive into what to do after a theft, check out [email protected] for complete guide on how to recover stolen crypto
-
12.06.26 17:13
keithwilson9899
ETHEREUM RECOVERY ASSISTANCE: CAPITAL CRYPTO RECOVER HELPED ME RECOVER $98,000 WORTH OF LOST ETH In cases of cryptocurrency scams, having accurate information and trusted support is essential. I would like to recommend Capital Crypto Recover Service, a professional team that specializes in assisting individuals with the recovery of lost or stolen Bitcoin and Ethereum (ETH). Their experienced experts are dedicated to helping victims of digital asset fraud by carefully analyzing each case, developing strategic recovery plans, Capital Crypto Recover Service knowledgeable team's primary goals are to satisfy clients and offer significant support and working diligently toward fund retrieval. The team is committed to providing reliable assistance and maintaining a high level of client satisfaction. Based on my assessment, their reputation professionalism and a strong commitment to their clients. If you have experienced a cryptocurrency loss, you can contacting them for further assistance Phone (Call/Text): +1 (336) 390-6684 Email: [email protected] Alternate Email: [email protected] Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
12.06.26 17:13
keithwilson9899
ETHEREUM RECOVERY ASSISTANCE: CAPITAL CRYPTO RECOVER HELPED ME RECOVER $98,000 WORTH OF LOST ETH In cases of cryptocurrency scams, having accurate information and trusted support is essential. I would like to recommend Capital Crypto Recover Service, a professional team that specializes in assisting individuals with the recovery of lost or stolen Bitcoin and Ethereum (ETH). Their experienced experts are dedicated to helping victims of digital asset fraud by carefully analyzing each case, developing strategic recovery plans, Capital Crypto Recover Service knowledgeable team's primary goals are to satisfy clients and offer significant support and working diligently toward fund retrieval. The team is committed to providing reliable assistance and maintaining a high level of client satisfaction. Based on my assessment, their reputation professionalism and a strong commitment to their clients. If you have experienced a cryptocurrency loss, you can contacting them for further assistance Phone (Call/Text): +1 (336) 390-6684 Email: [email protected] Alternate Email: [email protected] Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
14.06.26 14:53
Freeman James
Recently, I was scammed out of $332,000 in a fraudulent Bitcoin investment scheme. This devastating loss added significant stress to my already difficult health challenges, as I was also facing surgery expenses for cancer. Desperate to recover my funds, I spent countless hours researching and speaking with other victims. That effort led me to a Google post that revealed the excellent reputation of FundsRetriever. Only after many hours of digging and consulting others did I learn about their stellar track record. I decided to contact them because of their successful recovery history and encouraging client testimonials. I had no idea that this decision would become the turning point in my fight against cryptocurrency theft. Thanks to their expert team, I was able to recover my lost funds. The process was complex, but FundsRetriever's commitment to using the latest technology ensured a successful outcome. I highly recommend their services to anyone who has fallen victim to cryptocurrency fraud. For assistance, contact: 📧 [email protected] 📞 WhatsApp: +1 603 512 144 8, Telegram: @FundsRetriever
-
14.06.26 15:37
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
14.06.26 15:37
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
14.06.26 16:34
Emmi Hakola
I’m open about my experience with Bitcoin investment and losing money to scammers. That said, it is possible to recover stolen Bitcoin. I used to think recovery was impossible because that’s what I had been told. But last October, I fell for a forex scam promising extremely high returns and ended up losing nearly $87,600. After searching for help for a month, I came across a Reddit article about recovering stolen cryptocurrency. I reached out to the contact provided: [email protected] and WhatsApp +19852969146. I was scared and skeptical, having heard many bad stories, but I decided to give them a try. To my amazement, I got all my stolen Bitcoin back within a very short time. I’m not sure if I’m allowed to post links here, but you can reach out to them if you also need help.
-
14.06.26 16:53
James willson
I lost $328,650 to a fraudulent website that claimed to be a legitimate investment platform offering high returns. I was drawn in by the desire to earn more for myself and my family. Unfortunately, by the end of 2024, I realized it was a scam when the broker stopped responding to my emails and messages. A colleague then introduced me to ResQPro Firm, and to my surprise, they were able to trace and recover my stolen funds. Contact them at: resqprofirm AT AOL dot com | WhatsApp: +1 985 296 9146 | Telegram: ResQproFirm
-
14.06.26 19:38
riley777
G`DAY, I lost more than 119,000 Australian dollars to a crypto scam and it took almost everything I had saved which left me feeling like I had no future. I was stuck. I did not know where to go or how to find the money again. The wallet company is no help at all and they make it so hard to see where the coins go once they leave your account so you just feel lost. I spent days looking for a way out. Then I saw a post for a person who finds stolen money. The ad said they can track any crypto that goes missing. I wanted to check if it was real. I sent an email to [email protected] +44//// 7476618364\ to see if they could help me get my funds back. They did an amazing job. My money was back in my account in less than a week after they did a fast search and return.
-
15.06.26 06:12
Evan Garrison
When investing in staking platforms, proceed with caution. If your funds are stolen by a fake staking pool, the experience can be very frustrating. Rather than giving in to frustration, it's important to act quickly to improve your chances of recovering your money. Unfortunately, many victims never get their money back because scammers are often in another country or using fake identities. However, in some cases, tracking the funds is easier, especially for smart contract forensics specialists. I lost €18,500 to StakeKing. FundsRetriever found a backdoor in the contract and recovered my stake. Contact [email protected], WhatsApp +1(603)5121(448), or Telegram FUNDSRETRIEVER for assistance.
-
15.06.26 06:25
Glenn robble
Stop putting money into platforms promising guaranteed monthly returns of 10%, 20%, or more. These are Ponzi schemes. Your "profits" are just other victims' deposits. The moment withdrawals slow down, the scam is about to collapse. If you already have money trapped, do not send more to "unlock" your funds. That is a second scam. Instead, gather all transaction hashes and wallet addresses. Bitcoin Evolution Pro took €25,000 from me. FundsRetriever traced the funds through KYC exchanges and recovered my principal. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 06:34
Sallymarch
If IQ Option or any similar platform blocks your withdrawal citing "bonus terms" or "abnormal activity," do not argue with their chat support. They are not empowered to help you. Instead, request all trade logs and bonus terms in writing. Then get FundsRetrievers forensic specialist to audit your account. IQ Option held my €9,200 for two months. FundsRetriever reviewed my case, identified regulatory violations, and secured my full payout within 72 hours. Professional pressure works. Do it immediately. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 06:38
Sallymarch
If IQ Option or any similar platform blocks your withdrawal citing "bonus terms" or "abnormal activity," do not argue with their chat support. They are not empowered to help you. Instead, request all trade logs and bonus terms in writing. Then get FundsRetrievers forensic specialist to audit your account. IQ Option held my €9,200 for two months. FundsRetriever reviewed my case, identified regulatory violations, and secured my full payout within 72 hours. Professional pressure works. Do it immediately. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 06:41
Ewaguz
Cloud mining contracts are almost always too good to be true. I learned that the hard way with MineMax. First two months, small daily payouts. Then "maintenance fees" ate everything. Then my account was frozen. Then the website disappeared. I was heartbroken. FundsRetriever traced my payments through three shell companies to a real bank account. They froze it and got my €11,000 back. Recovery is possible even from complex scams. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 12:49
Jason
Did a scammer take your money? Fake loan, crypto fraud, romance trap, phishing—they count on you feeling helpless. Prove them wrong. FundsRetriever recovers stolen digital assets fast. No upfront payment. Ever. Backed by the FBI, Interpol, and cybercrime units. Blockchain tracing, legal freezing, and full recovery—for Bitcoin, Ethereum, USDT, Ponzi schemes, you name it. Your move: get a free case review right now. Then forensics, legal action, and your funds back. ⏳ Time is everything. 📧 FUNDSRETRIEVER @ PROTON.ME 📞 +16035121448 (WhatsApp) 📱 Telegram: @FUNDSRETRIEVER
-
15.06.26 12:56
Hillary
As a blockchain forensic analyst, I’ve reviewed numerous recovery cases. Fundsretriever demonstrates proper on-chain tracing, evidence preservation, and legal coordination. Their methodology helped several of my clients retrieve stolen or stuck assets. Recommended for victims seeking verifiable solutions. 📧 [email protected] Telegram @FUNDSRETRIEVER WhatsApp +1 603 512 1448
-
15.06.26 13:03
Feliksa Stegniy
A woman added me on Facebook, and after she suggested we become friends, we started communicating. Over time, she introduced me to a crypto trading platform called btctradingfx.com. She shared a lot of information about it, along with screenshots that made the platform seem trustworthy. Convinced by her claims, I decided to give it a try. I was promised a 10% weekly return, so I made an initial investment of $500. To my surprise, I received $5,000 back. That success encouraged me to invest more, so I put in $20,000. But when I tried to withdraw my funds, I was denied access and told I needed to deposit even more money before I could make a withdrawal. In the end, I lost a total of $43,850. It was an extremely difficult and painful experience. Fortunately, I later found a professional recovery service called ResQprofirm while searching on Google. I contacted them and provided all the evidence I had. They took my case seriously and were able to track down and recover my capital from the platform, which had been inaccessible for a long time. If you find yourself in a similar situation, you might consider reaching out to them via email at [email protected] or on WhatsApp at +19852969146, Telegram @resqprofirm Thank you, ResQPro, for your support.
-
15.06.26 13:05
James willson
The Most Credible Crypto Recovery Service: RESQPROFIRM RESQPROFIRM is a reliable, legitimate company that helps recover lost cryptocurrency assets. After weeks of doubting whether my lost BTC could ever be restored, I realized how widespread crypto scams have become. Caution is essential when dealing with strangers online, especially about money. While recovering stolen crypto is possible, avoiding fake "recovery companies" is just as important. Real hackers work discreetly and don't advertise openly. I was scammed multiple times while desperately seeking help. Finally, a friend introduced me to RESQPROFIRM—a trustworthy, discreet team. They handle everything from website security to crypto asset recovery. With their help, I recovered $320,000 in USDT within a week. Their professionalism, discretion, and speed were outstanding. If you've been compromised, don't lose hope—but beware of fraudsters posing as saviors. RESQPROFIRM are true professionals. I'm living proof. Contact them at [email protected], WhatsApp +19852969146, or Telegram @resqprofirm.
-
15.06.26 13:06
Tansy
Lost $18,500 to a fake Elon Musk crypto giveaway. Sent ETH, got nothing. Recovery pages demanded more gas fees. I stopped believing. FuNds rEtRiEveR on Te.le_gram was the real one. Email: [email protected] – WhatsApp: +1 603 512 1448
-
15.06.26 13:08
Sarahy billy
A REAL EXPERIENCE, EVERYONE ... PLEASE BE CAREFUL ONLINE A few weeks ago, I lost around $64,000 to a fake crypto trading platform. I was drawn in by the promise of earning 15% profit daily. It was a devastating time—I struggled to pay my bills and was financially ruined. I eventually opened up to a close friend, who recommended a crypto recovery team with highly effective methods. I contacted them, and they successfully recovered all my stolen digital assets with ease. Their service was excellent, and they acted quickly—within just 5 working days, they tracked down the scammers and returned my funds. I strongly urge anyone facing investment theft or similar issues to reach out to this team for the right solution and avoid losing large sums to fraudsters... Email: Resqprofirm @aol.com WhatsApp: +19852969146, telegram @resqprofirm
-
15.06.26 13:12
Cole donald
"I strongly recommend RESQPRO FIRM to anyone trying to recover lost cryptocurrency assets, including Bitcoin, USDC, USDT, Ethereum, and Trump Coin. Like many others, I was shocked to learn that crypto holdings can be stolen even when private keys are carefully protected. After a sophisticated hack wiped out my entire portfolio, I felt completely helpless. Fortunately, I was referred to RESQPRO FIRM. Their team understood the complexity of my situation and successfully recovered my funds. They were responsive, communicated clearly, and followed a careful, step-by-step process—which gave me a lot of reassurance during a stressful time. If you've experienced a similar financial loss, I encourage you to reach out to them. Their professionalism and ethical hacking skills exceeded my expectations." Contact Info: · WhatsApp: +1 (985) 2969146 · Email: [email protected] · Telegram: Resqprofirm
-
15.06.26 13:16
Meral Yetkiner
I recently lost $38,000 to an online platform. Initially, they requested additional deposits to grant me access to my portfolio. Despite complying, my withdrawal requests were repeatedly denied, and they continued asking for more funds. Suspecting fraudulent activity, I ceased further payments and promptly reported the matter to ResQProfirm, a firm I discovered through Google. They listened to my situation, initiated communication regarding the sequence of events, and requested all relevant evidence to support their investigation. Through their dedicated efforts, they successfully traced and recovered my funds. I extend my thanks to ResQProfirm at [email protected] and via WhatsApp at +19852969146. I urge everyone to exercise caution and thoroughly research any platform before investing.
-
15.06.26 13:18
Silas Olsen
A fraudulent investment scheme operated by BTCMining.limited functions as a fake return scam. In this setup, scammers lure victims with false promises of high returns. Through manipulative tactics, they gain individuals' trust and convince them to invest, ultimately leading to financial loss. If you have ever faced a cyber threat or fallen victim to an online crypto scam and need to reach the authorities, I recommend contacting [email protected]. They are a legitimate team that helps victims of online crypto scams using advanced tools.
-
15.06.26 13:59
Ewaguz
If a binary options broker refuses your withdrawal, do not pay any "verification fees" or "tax fees." These are lies designed to extract more money. Stop communicating with their support team – they are trained to stall. Instead, immediately document every transaction, screenshot your account balance, and contact a professional recovery specialist. BinaryBook stole €14,500 from me before I learned this. FundsRetriever traced the deposits and recovered everything within two weeks. Do not wait. Do not pay more fees. Act now. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:16
Martina k.
Stop putting money into platforms promising guaranteed monthly returns of 10%, 20%, or more. These are Ponzi schemes. Your "profits" are just other victims' deposits. The moment withdrawals slow down, the scam is about to collapse. If you already have money trapped, do not send more to "unlock" your funds. That is a second scam. Instead, gather all transaction hashes and wallet addresses. Bitcoin Evolution Pro took €25,000 from me. FundsRetriever traced the funds through KYC exchanges and recovered my principal. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:18
Garrison Good
If IQ Option or any similar platform blocks your withdrawal citing "bonus terms" or "abnormal activity," do not argue with their chat support. They are not empowered to help you. Instead, request all trade logs and bonus terms in writing. Then hire a forensic specialist to audit your account. IQ Option held my €9,200 for two months. FundsRetriever reviewed my case, identified regulatory violations, and secured my full payout within 72 hours. Professional pressure works. Do it immediately. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:22
Sallymarch
Never grant API keys with withdrawal permissions to any third-party software. This is how crypto arbitrage bots steal your funds. If you have already done this, revoke all API keys immediately. Then check your exchange transaction history. CryptoArb AI drained €7,800 from my account within hours. FundsRetriever reverse-engineered the bot's code, traced the scammer's wallet, and recovered everything. Always use "read-only" API permissions only. If you made the mistake, act fast. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:23
Glennrobble
If a binary options broker closes your account and confiscates your profits, do not accept their explanation. Demand a full audit of your trade history. Most brokers cannot justify their actions when challenged by professionals. ExpertOption stole €6,200 from me claiming "abnormal activity." FundsRetriever audited my trades, proved they were legitimate, and threatened legal action. The broker paid within 10 days. Do not let them intimidate you. Get professional help. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:25
Evan Garrison
Cloud mining contracts are almost always too good to be true. I learned that the hard way with MineMax. First two months, small daily payouts. Then "maintenance fees" ate everything. Then my account was frozen. Then the website disappeared. I was heartbroken. FundsRetriever traced my payments through three shell companies to a real bank account. They froze it and got my €11,000 back. Recovery is possible even from complex scams. Contact [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 14:26
Ewaguz
That 100% deposit bonus looks tempting, doesn't it? I took it. Big mistake. When I tried to withdraw my €4,500, Olymp Trade demanded I trade 50 times the bonus amount. Impossible by design. My money was trapped. FundsRetriever reviewed the terms and found they violated consumer protection laws in my country. They negotiated directly with Olymp Trade's legal team. Within a week, my funds were released. My advice? Never accept bonuses. But if you're already trapped, call [email protected], WhatsApp +1(603)5121(448) or Telegram FUNDSRETRIEVER.
-
15.06.26 16:34
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
15.06.26 16:34
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
15.06.26 16:41
Louane Mercier
It is crucial to act quickly and consult a reputable, experienced recovery specialist who will support you throughout the entire recovery process. You must provide them with transaction evidence, scammer information, and any other relevant details that could aid the investigation. With this data, the experts can trace and attempt to recover your funds from the scammers' concealed accounts or wallets. R£sQprofirm company offers recovery assistance with no upfront fees. Contact them via Telegram (@ResQprofirm), WhatsApp (+19852969146), or email ([email protected]).
-
15.06.26 16:45
Andrés Montero
I’m open about my experience with Bitcoin investment and losing money to scammers. That said, it is possible to recover stolen Bitcoin. I used to think recovery was impossible because that’s what I had been told. But last October, I fell for a forex scam promising extremely high returns and ended up losing nearly $87,600. After searching for help for a month, I came across a Reddit article about recovering stolen cryptocurrency. I reached out to the contact provided: [email protected] and WhatsApp +19852969146. I was scared and skeptical, having heard many bad stories, but I decided to give them a try. To my amazement, I got all my stolen Bitcoin back within a very short time. I’m not sure if I’m allowed to post links here, but you can reach out to them if you also need help.
-
15.06.26 16:48
Olivia Sørensen
Several months ago, investing in Bitcoin proved to be one of my most lucrative endeavors. I achieved considerable profits across multiple platforms and felt a strong sense of accomplishment. Unfortunately, the situation deteriorated when I inadvertently engaged with a fraudulent Bitcoin platform. This entity swindled me out of $92,000 USD, refused to honor my withdrawal requests, and persistently demanded further deposits. Fortunately, I encountered (R£SQPRO FIRM) online. After reporting my case to them, they acted promptly and effectively recovered my lost Bitcoin. I am sincerely grateful for their professionalism and continuous assistance. Contact: ResQprofirm AT aol.com, Telegram @resqprofirm, WhatsApp +1 9 8 5 2 9 6 9 1 4 6.
-
15.06.26 16:51
Viljar Yohannes
I'm willing to share my experience with Bitcoin investment and losing money to scammers. But yes, recovering stolen Bitcoin is possible. I never believed in Bitcoin recovery myself, because I was told it couldn't be done. Then, last October, I fell for a forex scam that promised unrealistically high returns, and I ended up losing nearly $70,000. I searched for help for about a month until I finally found a Reddit article about recovering stolen cryptocurrency. I reached out to the contact mentioned: [RESQPROFIRM [at] AOL DOT com] and [WhatsApp +19852969146]. I was scared and skeptical because I'd heard horror stories, but I decided to give them a try. To my surprise, I got all my stolen Bitcoin back from the scammers in a very short time. I'm not sure if I'm allowed to post links here, but you can contact them if you need help too.
-
15.06.26 16:58
Guimar da Rosa
Withdrawal troubles shouldn’t stress you out. I faced a similar problem, and this firm stepped in and recovered my funds. Their support truly mattered. Contact them: [ResQProFirm @aol.com] telegram @resqprofirm, WhatsApp: <+198> <5296> <9146>.
-
15.06.26 17:03
Andrea Escalante
If withdrawals keep getting denied, stay calm. I went through the same, and this firm helped me recover everything. Their assistance was outstanding. Contact: [[email protected]], Telegram: ResQprofirm, WhatsApp: <+198> <5296> <9146>. Withdrawal troubles shouldn’t
-
16.06.26 11:40
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
16.06.26 11:43
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
16.06.26 13:37
Felix Steve
MY CRYPTO WAS STOLEN – HERE'S HOW I GOT IT BACK I'm Felix Steve from Canada, and I lost $115,000 USDC to a fraudulent broker who locked me out of my wallet. After sleepless nights, a friend told me about RESQPROFIRM Recovery Service. I sent them my wallet addresses, transaction history, and chat logs. Their team used blockchain tracking to trace the stolen funds, identified the scammer's wallet, and froze the assets before they could be moved. Within 24 hours, most of my crypto was recovered. I can't thank them enough. If you need help, reach out via WhatsApp: +19852969146, email: [email protected], or TG: @resqprofirm.
-
16.06.26 13:45
Wills ben
SUCCESSFUL CRYPTO SCAM RECOVERY – HOW I REGAINED ACCESS TO MY LOST WALLET My name is Felix Steve, and I'm from Canada. I'm sharing my story to help others who have fallen victim to crypto fraud. A few months ago, I was lured into a fake investment scheme promoted by a broker company. With Bitcoin prices climbing, I invested heavily—only to lose $115,000 USDC when the broker locked me out of my wallet and assets. It was a harrowing experience that left me sleepless and desperate. Crypto scams are on the rise, often involving bogus trading platforms, phishing, and misleading promises. In my search for help, a fellow crypto enthusiast recommended RESQPROFIRM Recovery Service, which specializes in recovering lost or stolen funds. After checking their reviews, I reached out and supplied all the evidence—wallet addresses, transaction records, and communication logs. Their team responded immediately and launched an investigation. Using advanced blockchain tracking, they traced the stolen funds, pinpointed the scammer's wallet, and worked with authorities to freeze the assets in time. Remarkably, within just 24 hours, RESQPROFIRM recovered the bulk of my stolen crypto. I was overwhelmed with relief and gratitude. Their professionalism, transparency, and steady communication made all the difference during a very dark period. If you've been scammed, I wholeheartedly recommend contacting them via WhatsApp: +19852969146, email: [email protected], or Telegram: @resqprofirm.
-
18.06.26 13:31
Noemi Bernard
I never expected such outstanding results. The outcome far exceeded my expectations, and I am extremely satisfied with the successful recovery of my stolen funds totaling $49,360 from my blockchain wallet. I hold this team in the highest regard. Without a doubt, they are among the most dedicated professionals in the field of fund recovery. Keep up the exceptional work! Email: [email protected] WhatsApp: +1 985 296 9146
-
18.06.26 13:35
Carter Morris
My experience improved significantly thanks to ResQprofirm's expert assistance and attentive customer care. Their professionalism was evident every step of the way they were able to track and recover my stolen crypto $88,360, email: [email protected], WhatsApp +19852969146.
-
18.06.26 13:40
Kuybida Andriyiv
I recovered my $232,000 refund through the assistance of [email protected] and WhatsApp +19852969146. Their guidance was very helpful.
-
20.06.26 14:57
michaeldavenport218
I was recently scammed out of $53,000 by a fraudulent Bitcoin investment scheme, which added significant stress to my already difficult health issues, as I was also facing cancer surgery expenses. Desperate to recover my funds, I spent hours researching and consulting other victims, which led me to discover the excellent reputation of Capital Crypto Recover, I came across a Google post It was only after spending many hours researching and asking other victims for advice that I discovered Capital Crypto Recovery’s stellar reputation. I decided to contact them because of their successful recovery record and encouraging client testimonials. I had no idea that this would be the pivotal moment in my fight against cryptocurrency theft. Thanks to their expert team, I was able to recover my lost cryptocurrency back. The process was intricate, but Capital Crypto Recovery's commitment to utilizing the latest technology ensured a successful outcome. I highly recommend their services to anyone who has fallen victim to cryptocurrency fraud. For assistance contact [email protected] and on Telegram OR Call Number +1 (336)390-6684 via email: [email protected] you can visit his website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
20.06.26 14:57
michaeldavenport218
I was recently scammed out of $53,000 by a fraudulent Bitcoin investment scheme, which added significant stress to my already difficult health issues, as I was also facing cancer surgery expenses. Desperate to recover my funds, I spent hours researching and consulting other victims, which led me to discover the excellent reputation of Capital Crypto Recover, I came across a Google post It was only after spending many hours researching and asking other victims for advice that I discovered Capital Crypto Recovery’s stellar reputation. I decided to contact them because of their successful recovery record and encouraging client testimonials. I had no idea that this would be the pivotal moment in my fight against cryptocurrency theft. Thanks to their expert team, I was able to recover my lost cryptocurrency back. The process was intricate, but Capital Crypto Recovery's commitment to utilizing the latest technology ensured a successful outcome. I highly recommend their services to anyone who has fallen victim to cryptocurrency fraud. For assistance contact [email protected] and on Telegram OR Call Number +1 (336)390-6684 via email: [email protected] you can visit his website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
21.06.26 11:09
Maurizio Rolland
I would like to express my sincere appreciation to RESQPRO FIRM for their outstanding assistance in helping victims of online fraud. Many scammers deceive investors by blocking withdrawals and continuously demanding additional deposits, making the loss of hard-earned funds a painful experience. Fortunately, RESQPRO FIRM provides support to individuals seeking to recover funds lost to fraudulent online schemes. Contact: Email: RESQPRO FIRM at Gmail Telegram: RESQPROFIRM, [email protected], WhatsApp: +1 985 296 9146
-
21.06.26 11:13
Buse Fahri
It is important for more people to stand together in the fight against online fraud. Those who target innocent individuals especially vulnerable people such as seniors should be held fully accountable for their actions. Every effort to raise awareness and support victims makes a meaningful difference. The team at RESQPRO FIRM is committed to helping expose fraudulent schemes and assisting those affected by online scams. Their dedication, persistence, and passion for protecting victims are truly commendable. I sincerely appreciate the hard work and commitment shown toward this mission. Together, we can continue to educate others, support victims, and work toward a safer online environment for everyone. Contact Information: Telegram: RESQPROFIRM WhatsApp: +1 985 296 9146 Email: [email protected], [email protected]
-
21.06.26 11:16
علیرضا گلشن
The successful recovery of my stolen funds, totaling $1,310,000, would not have been possible without your unwavering support, dedication, and tireless efforts. I am truly grateful for the opportunity to work with such a skilled and professional team. From the very beginning, I had confidence in your ability to handle this challenging situation, and you exceeded my expectations by delivering remarkable results. Your expertise, persistence, and commitment throughout the process were exceptional. I encourage you to continue maintaining the high standards of professionalism and excellence that distinguish your work. You exemplify the qualities of a trustworthy, dedicated, and hardworking professional, and your efforts deserve sincere recognition and appreciation. Contact: Email: [email protected], [email protected], Telegram: Resqprofirm WhatsApp: +1 985 296 9146
-
22.06.26 21:51
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
22.06.26 21:51
kimberlyhebert786
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
24.06.26 01:25
Fraddy Pual
I never thought it would happen to me—but I lost $256,100 in Bitcoin through a shady investment deal. I was shattered, panicked, and convinced my savings were gone forever. Just when I was about to give up, I stumbled upon reviews for FUNDSRETRIEVER, a cyber recovery team with a solid reputation. I decided to give it a shot, and to my absolute shock, they recovered every single cent in record time. Working with them was a lifesaver. If you've been tricked by fake investment platforms, don't lose hope—FUNDSRETRIEVER can help. Contact them here: Email: [email protected] | WhatsApp: +1603512144 8| Telegram: @Fundsretriever
-
24.06.26 01:27
Fraddy Pual
I never thought it would happen to me—but I lost $256,100 in Bitcoin through a shady investment deal. I was shattered, panicked, and convinced my savings were gone forever. Just when I was about to give up, I stumbled upon reviews for FUNDSRETRIEVER, a cyber recovery team with a solid reputation. I decided to give it a shot, and to my absolute shock, they recovered every single cent in record time. Working with them was a lifesaver. If you've been tricked by fake investment platforms, don't lose hope—FUNDSRETRIEVER can help. Contact them here: Email: [email protected] | WhatsApp: +16035121448 | Telegram: @Fundsretriever
-
24.06.26 01:28
Fraddy Pual
I never thought it would happen to me—but I lost $256,100 in Bitcoin through a shady investment deal. I was shattered, panicked, and convinced my savings were gone forever. Just when I was about to give up, I stumbled upon reviews for FUNDSRETRIEVER, a cyber recovery team with a solid reputation. I decided to give it a shot, and to my absolute shock, they recovered every single cent in record time. Working with them was a lifesaver. If you've been tricked by fake investment platforms, don't lose hope—FUNDSRETRIEVER can help. Contact them here: Email: [email protected] | WhatsApp: +16035121448 | Telegram: @Fundsretriever.
-
24.06.26 01:58
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
24.06.26 01:58
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
24.06.26 14:16
Universina da Mota
Becoming a victim of an investment scam is never anyone's intention it often happens because fraudsters exploit trust and a lack of awareness. I would like to express my sincere gratitude to the dedicated team at ResQpro for their professionalism and commitment to helping victims of online investment fraud. Their efforts in assisting individuals with the recovery of stolen assets and holding scammers accountable are truly commendable. If you need assistance or would like to learn more, you can contact them through: Email: [email protected] Alternative Email: [email protected] Telegram: @ResQprofirm WhatsApp: +1 (985) 296-9146
-
24.06.26 14:21
Elizabeth Thompson
If you believe you have been the victim of an investment scam, it is important to act promptly and gather all relevant information. Keep records of transaction receipts, wallet addresses, communication logs, account details, and any other evidence related to the incident. Providing accurate documentation can help investigators, financial institutions, legal professionals, or recovery specialists review your case and determine what options may be available. Be cautious of anyone who guarantees the recovery of lost funds or requests large upfront payments. For additional information, you may contact: Email: [email protected] Telegram: @ResQprofirm WhatsApp: +1 (985) 296-9146
-
24.06.26 15:33
Júlia Castro
If you have fallen victim to an investment scam, it is important to act quickly and gather all available evidence related to the incident. This may include transaction records, wallet addresses, screenshots of conversations, emails, account details, and any information connected to the individuals or entities involved. Having complete documentation can help professionals assess your situation and explore possible recovery options. Always exercise caution when seeking assistance and carefully verify the credentials of any service provider before proceeding. For further information, you may contact: Email: [email protected] Telegram: @ResQprofirm WhatsApp: +1 (985) 296-9146
-
24.06.26 22:01
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
24.06.26 22:01
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
25.06.26 21:13
Emilie Safi
A fraudulent investment scheme operated by BTCMining.limited functions as a fake return scam. In this setup, scammers lure victims with false promises of high returns. Through manipulative tactics, they gain individuals' trust and convince them to invest, ultimately leading to financial loss. If you have ever faced a cyber threat or fallen victim to an online crypto scam and need to reach the authorities, I recommend contacting [email protected], [email protected], WhatsApp +19852969146, telegram @resqprofirm. They are a legitimate team that helps victims of online crypto scams using advanced tools.
-
25.06.26 21:25
Emilie Safi
So I ended up losing $38,000 to this platform. At first, they kept asking me to put in more money so I could get into my portfolio. I did that, but then they wouldn’t let me withdraw anything—just kept asking for more deposits. It got way too suspicious, so I stopped. I found this company called ResQProfirm on Google and told them what happened. They got in touch, asked me to walk them through everything, and I gave them all the proof I had. They did an amazing job tracking down my money and getting it back. Big thanks to them at [email protected] and on WhatsApp at +19852969146. Please be careful out there and always research before investing.
-
26.06.26 01:04
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
26.06.26 01:04
robertalfred175
CRYPTO SCAM RECOVERY SUCCESSFUL – A TESTIMONIAL OF LOST PASSWORD TO YOUR DIGITAL WALLET BACK. My name is Robert Alfred, Am from Australia. I’m sharing my experience in the hope that it helps others who have been victims of crypto scams. A few months ago, I fell victim to a fraudulent crypto investment scheme linked to a broker company. I had invested heavily during a time when Bitcoin prices were rising, thinking it was a good opportunity. Unfortunately, I was scammed out of $120,000 AUD and the broker denied me access to my digital wallet and assets. It was a devastating experience that caused many sleepless nights. Crypto scams are increasingly common and often involve fake trading platforms, phishing attacks, and misleading investment opportunities. In my desperation, a friend from the crypto community recommended Capital Crypto Recovery Service, known for helping victims recover lost or stolen funds. After doing some research and reading multiple positive reviews, I reached out to Capital Crypto Recovery. I provided all the necessary information—wallet addresses, transaction history, and communication logs. Their expert team responded immediately and began investigating. Using advanced blockchain tracking techniques, they were able to trace the stolen Dogecoin, identify the scammer’s wallet, and coordinate with relevant authorities to freeze the funds before they could be moved. Incredibly, within 24 hours, Capital Crypto Recovery successfully recovered the majority of my stolen crypto assets. I was beyond relieved and truly grateful. Their professionalism, transparency, and constant communication throughout the process gave me hope during a very difficult time. If you’ve been a victim of a crypto scam, I highly recommend them with full confidence contacting: Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text: +1 (336) 390-6684 Website: https://recovercapital.wixsite.com/capital-crypto-rec-1
-
26.06.26 02:48
Miriam Rocha
I trusted this platform with $120,000 of my hard-earned money. Then they started asking for more deposits just so I could access my own portfolio. I paid, but every withdrawal request was denied. They kept pushing for more money. I finally stopped it just felt wrong. Desperate, I found ResQProfirm on Google. They didn't just hear me out; they truly listened. I shared all my proof, and they launched an investigation. Thanks to their hard work, they tracked and returned my funds. From the bottom of my heart, thank you to [email protected] and their WhatsApp +19852969146. Please stay safe and always verify a platform before investing
-
26.06.26 02:52
Miško Bakić
I got my $232,000 refund thanks to [email protected] and WhatsApp +19852969146. Highly recommended for anyone in a similar situation.
-
26.06.26 02:56
Asunción Herrera
A recovery of $48,330 was facilitated by [email protected]. Individuals who have experienced financial fraud may consider contacting this service.
-
26.06.26 15:05
Riley Stephens
If withdrawals keep getting denied, stay calm. I went through the same, and this firm helped me recover everything. Their assistance was outstanding. Contact: [ResQProFirm @Gmail|•|com], Telegram: ResQprofirm, WhatsApp: <+198> <5296> <9146>.
-
26.06.26 15:09
Antonio Riley
Withdrawal troubles shouldn’t stress you out. I faced a similar problem, and this firm stepped in and recovered my funds. Their support truly mattered. Contact them: [[email protected], ResQprofirm @aol.com], Telegram: ResQprofirm, WhatsApp: +19852969146.
-
00:37
kimberlyhebertt673
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
-
00:37
kimberlyhebertt673
I invested in bitcoin trading After losing $78.4 USDT) linked to a romance fraud scam worth of cryptocurrency through an online investment platform and later discovered it was a scam. After extensive research for recovery options, I contacted CAPITAL CRYPTO RECOVER based on positive client reviews and recommendations. Their professional security team guided me through the recovery process using advanced technology, and I was able to recover my lost cryptocurrency successfully. I am truly grateful for their support and assistance during such a difficult experience. I will advise you to contact CAPITAL CRYPTO RECOVER helped me recover my funds. For anyone facing similar issues, Website: https://recovercapital.wixsite.com/capital-crypto-rec-1 Email: [email protected] Telegram: @Capitalcryptorecover Contact: [email protected] Call/Text Number: +1 (336) 390-6684
Для участия в Чате вам необходим бесплатный аккаунт pro-blockchain.com
Войти
Регистрация
Похожие новости
Подпишись на telegram канал и узнавай первым обо всех новостях