Stray2024 commited on
Commit
d2cae38
·
verified ·
1 Parent(s): 0e034d7

Add bilingual Dataset Card with AI generation disclaimer

Browse files
Files changed (1) hide show
  1. README.md +77 -0
README.md ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - ru
4
+ - en
5
+ license: apache-2.0
6
+ task_categories:
7
+ - text-generation
8
+ tags:
9
+ - code
10
+ - python
11
+ - synthetic
12
+ - compiler-verified
13
+ - instruction-tuning
14
+ - unsloth
15
+ size_categories:
16
+ - n<1K
17
+ ---
18
+
19
+ # 🤖 Stray Python Verified Dataset
20
+
21
+ [English](#english) | [Русский](#русский)
22
+
23
+ ---
24
+
25
+ ## English
26
+
27
+ ### Overview
28
+ **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).
29
+
30
+ 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:
31
+ 1. **AST Parsing (`ast.parse`)**: Verifies syntactic accuracy and structural integrity.
32
+ 2. **Bytecode Compilation (`compile`)**: Validates CPython bytecode generation.
33
+ 3. **Execution Sandbox (`subprocess.run`)**: Verifies runtime behavior and filters out hanging loops (`timeout > 3s`) and runtime exceptions.
34
+
35
+ ### Dataset Format
36
+ The dataset is formatted in standard **ChatML (`messages`)** format:
37
+ ```json
38
+ {
39
+ "messages": [
40
+ {"role": "system", "content": "Ты опытный Python разработчик..."},
41
+ {"role": "user", "content": "<Task condition>"},
42
+ {"role": "assistant", "content": "```python\n<Verified Code>\n```"}
43
+ ]
44
+ }
45
+ ```
46
+
47
+ ### ⚠️ Disclaimer & Limitations
48
+ * This dataset was **synthetically generated by AI models** and filtered automatically through code compilers and execution sandboxes.
49
+ * 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**.
50
+ * Intended primarily for research, distillation, and fine-tuning experiments.
51
+
52
+ ---
53
+
54
+ ## Русский
55
+
56
+ ### Описание
57
+ **Stray Python Verified** — это очищенный синтетический датасет для дообучения (SFT / LoRA) компактных языковых моделей по программированию на Python (Qwen 2.5 Coder, Gemma, Llama).
58
+
59
+ Каждый пример создан двух-агентной фабрикой (**Dual-AI**) и прошёл автоматический **3-ступенчатый фильтр компилятора и рантайм-песочницы**:
60
+ 1. **AST-валидация**: отсекает синтаксические ошибки, кривые отступы и незакрытые скобки.
61
+ 2. **Компиляция байткода**: гарантирует корректность внутренней структуры Python C-байткода.
62
+ 3. **Изолированный запуск**: отлавливает зависающие циклы (`timeout > 3s`) и рантайм-падения.
63
+
64
+ ### Формат данных
65
+ Данные упакованы в стандартный формат **ChatML (`messages`)**, полностью готовый для обучения через **Unsloth, TRL, SFTTrainer** и Hugging Face:
66
+ * `system`: инструкция для роли разработчика.
67
+ * `user`: условие задачи по Python.
68
+ * `assistant`: протестированный рабочий блок кода.
69
+
70
+ ### ⚠️ Предупреждение и ограничения
71
+ * Датасет **сгенерирован искусственным интеллектом** и проверен компилятором.
72
+ * Несмотря на автоматическую проверку синтаксиса и выполнения, **в коде могут присутствовать логические недочёты, шероховатости или неоптимальные краевые случаи**.
73
+ * Создан для исследовательских целей, экспериментов по дистилляции знаний и дообучению легковесных моделей.
74
+
75
+ ---
76
+ **Author / Создатель:** [Stray2024](https://huggingface.co/Stray2024)
77
+ **Current Samples / Всего сэмплов:** 187