Stray2024's picture
Add bilingual Dataset Card with AI generation disclaimer
d2cae38 verified
|
Raw History Blame Contribute Delete
4.21 kB
metadata
language:
  - ru
  - en
license: apache-2.0
task_categories:
  - text-generation
tags:
  - code
  - python
  - synthetic
  - compiler-verified
  - instruction-tuning
  - unsloth
size_categories:
  - n<1K

🤖 Stray Python Verified Dataset

English | Русский


English

Overview

Stray Python Verified is a high-purity synthetic dataset designed for instruction-tuning and fine-tuning lightweight code LLMs (e.g., Qwen 2.5 Coder 0.5B/1.5B/3B, Gemma 2/3, Llama 3.2).

Each sample in this dataset was generated using a Dual-AI pipeline and subjected to a strict 3-stage automated compiler and sandbox verification before being accepted:

  1. AST Parsing (ast.parse): Verifies syntactic accuracy and structural integrity.
  2. Bytecode Compilation (compile): Validates CPython bytecode generation.
  3. Execution Sandbox (subprocess.run): Verifies runtime behavior and filters out hanging loops (timeout > 3s) and runtime exceptions.

Dataset Format

The dataset is formatted in standard ChatML (messages) format:

{
  "messages": [
    {"role": "system", "content": "Ты опытный Python разработчик..."},
    {"role": "user", "content": "<Task condition>"},
    {"role": "assistant", "content": "```python\n<Verified Code>\n```"}
  ]
}

⚠️ Disclaimer & Limitations

  • This dataset was synthetically generated by AI models and filtered automatically through code compilers and execution sandboxes.
  • While the code is guaranteed to be syntactically valid and non-crashing in basic environments, some subtle logical oversights, edge-case limitations, or imperfect implementations may still be present.
  • Intended primarily for research, distillation, and fine-tuning experiments.

Русский

Описание

Stray Python Verified — это очищенный синтетический датасет для дообучения (SFT / LoRA) компактных языковых моделей по программированию на Python (Qwen 2.5 Coder, Gemma, Llama).

Каждый пример создан двух-агентной фабрикой (Dual-AI) и прошёл автоматический 3-ступенчатый фильтр компилятора и рантайм-песочницы:

  1. AST-валидация: отсекает синтаксические ошибки, кривые отступы и незакрытые скобки.
  2. Компиляция байткода: гарантирует корректность внутренней структуры Python C-байткода.
  3. Изолированный запуск: отлавливает зависающие циклы (timeout > 3s) и рантайм-падения.

Формат данных

Данные упакованы в стандартный формат ChatML (messages), полностью готовый для обучения через Unsloth, TRL, SFTTrainer и Hugging Face:

  • system: инструкция для роли разработчика.
  • user: условие задачи по Python.
  • assistant: протестированный рабочий блок кода.

⚠️ Предупреждение и ограничения

  • Датасет сгенерирован искусственным интеллектом и проверен компилятором.
  • Несмотря на автоматическую проверку синтаксиса и выполнения, в коде могут присутствовать логические недочёты, шероховатости или неоптимальные краевые случаи.
  • Создан для исследовательских целей, экспериментов по дистилляции знаний и дообучению легковесных моделей.

Author / Создатель: Stray2024 Current Samples / Всего сэмплов: 187