Free Use: You may use this dataset for any purpose, including commercial, academic, or personal projects, free of charge. 2. Mandatory Attribution: If you use PyMAX in your research, to train a model, or in any publication, you MUST explicitly mention "PyMAX by Inserloft Research (Research.Inserloft.com)" in the System Card, Model Card, or documentation where you cite public datasets used.

The largest synthetic Python dataset for training and fine-tuning LLMs.

📖 Overview

PyMAX is a massive, high-quality synthetic dataset of 500 million Python code examples created by Inserloft Research. It is specifically designed to train and fine-tune Large Language Models (LLMs) specialized in programming, code generation, and software development tasks.

This dataset bridges the gap between general-purpose models and specialized coding models, providing an unprecedented volume of Python-specific training data in a clean, structured, and ready-to-use format. PyMAX is not just a collection of prompts and responses; it is a foundational-grade dataset enriched with technical metadata, syntactic validation, and conversational alignment for modern LLMs.

🔥 Key Features

Feature	Description
Scale	500,000,000 examples (100 Parquet shards)
Languages	English (50%) & Spanish (50%)
Format	Parquet (Snappy compression), optimized for streaming
Size	~60 GB compressed
Generator	Enhanced Python script leveraging multiprocessing and AST analysis
Topics	Basic syntax, OOP, data structures, algorithms, web dev, snippets, etc.
Structure	Prompt, Response, Code, Explanation, Metadata, Messages, Test Code, Safety Flags
Validation	Automatic syntactic validation via `ast.parse`
Metadata	Complexity scores, imports, docstring detection, algorithmic complexity, tags
Quality	100% synthetic generation with rigorous filtering and deduplication
Use Cases	Code LLMs, fine-tuning, instruction tuning, RLHF, evaluation

🎯 Use Cases

Fine-tuning Code LLMs: Train models like CodeLlama, StarCoder, or custom architectures.
Instruction Tuning: Build models that understand natural language programming instructions.
Benchmarking: Evaluate coding capability of LLMs.
Education: Create AI tutors for Python learning.
Research: Study code generation, program synthesis, and software engineering.
Execution-Based Evaluation: Leverage included test cases for pass/fail scoring.
Safety-Aware Training: Filter examples using the built-in is_safe flag.

📊 Dataset Structure

Each example in PyMAX contains the following fields:

Field	Type	Description
`prompt`	`string`	A question or instruction about Python (English or Spanish)
`response`	`string`	Detailed explanation followed by code block
`code`	`string`	The raw source code
`explanation`	`string`	Textual explanation of the code
`messages`	`list`	Conversational format (ChatML): `[{"role": "user", "content": prompt}, {"role": "assistant", "content": response}]`
`test_code`	`string`	Validation code (assertions) for execution-based checking
`metadata`	`dict`	Contains: `topic`, `subtopic`, `difficulty`, `language`, `python_version`
`syntax_valid`	`bool`	Whether the code passes `ast.parse` (always `True` after filtering)
`imports`	`list`	List of all imported libraries extracted from the AST
`has_docstring`	`bool`	Whether the main module/function/class includes a docstring
`complexity_score`	`int`	Sum of lines of code + number of AST nodes
`hash`	`string`	SHA-256 hash of `prompt + response` for deduplication
`time_complexity`	`string`	Algorithmic time complexity (e.g., "O(N)")
`space_complexity`	`string`	Algorithmic space complexity (e.g., "O(1)")
`code_length`	`int`	Number of characters in the `code` field
`tags`	`list`	Array combining topic, subtopic, and difficulty (e.g., `["data_structures", "lists", "beginner"]`)
`is_safe`	`bool`	Whether the code avoids dangerous calls (`eval`, `exec`, `os.system`, `subprocess`)

🗂️ Topics Covered

Basic Syntax: Variables, operators, data types.
Control Structures: Conditionals, loops, comprehensions.
Functions: Definitions, decorators, generators, closures.
OOP: Classes, inheritance, polymorphism, magic methods.
Data Structures: Lists, dicts, sets, tuples, custom structures.
Algorithms: Sorting, searching, dynamic programming.
File I/O: Reading/writing files, serialization.
Web Development: Flask, Django, FastAPI basics.
Testing: Unit tests, mocking, pytest.
Concurrency: Threading, asyncio, multiprocessing.
Snippets: Short modular code blocks (utilities, math ops, list manipulations) for focused training.

🚀 How to Use

Installation

# Install datasets library
pip install datasets

# Load PyMAX (requires access approval)
from datasets import load_dataset

dataset = load_dataset("Inserloft/PyMAX", split="train")

Access a Single Example

example = dataset[0]
print(example["prompt"])
print(example["code"])
print(example["messages"])       # Conversational format
print(example["test_code"])      # Assertions for validation
print(example["time_complexity"])# "O(N)"

Stream Large Batches

from datasets import load_dataset

dataset = load_dataset("Inserloft/PyMAX", split="train", streaming=True)
for batch in dataset.iter(batch_size=1000):
    process(batch)

📋 Technical Details

Metric	Value
Total Examples	500,000,000
Languages	English (50%), Spanish (50%)
Format	Parquet (Snappy compression)
Compressed Size	~60 GB
Number of Shards	100 (each ~5 million examples)
Generation Time	~1.3 hours (4770 seconds) on 48-core infrastructure
Parallelism	`multiprocessing.Pool` with 48 workers
Batch Size	5,000,000 rows per batch (avoids memory overflow)
Syntax Validation	Automated via `ast.parse`
Deduplication	SHA-256 hash of `prompt+response`
Safety Check	Flags dangerous calls (`eval`, `exec`, `os.system`, `subprocess`)
Python Version Compatibility	3.10+

⚙️ Dataset Generation Pipeline

PyMAX was generated using an enhanced Python script by NaNo 3.4 (Inserloft's proprietary AI model). The pipeline includes:

Extreme Parallel Processing: Utilizes multiprocessing.Pool with 48 workers, fully leveraging the available hardware.
Strategic Batching: Data is generated and packaged in stable chunks of 5 million rows per batch, resulting in exactly 100 Parquet files. This prevents OOM errors and enables efficient streaming.
AST Analysis: Each code snippet is parsed using Python's ast module to:
- Validate syntax (syntax_valid).
- Extract imports (imports).
- Detect docstrings (has_docstring).
- Compute complexity score (lines + AST nodes).
Metadata Enrichment: Algorithmic complexity (time_complexity, space_complexity), code length, and unified tags are automatically calculated.
Conversational Alignment: Each example is converted into a ChatML-compatible messages list, making the dataset instantly ready for instruction tuning.
Test Case Injection: Each example includes a test_code snippet with assertions, enabling execution-based evaluation and potential RLHF workflows.
Safety Filtering: A scanner flags examples containing eval, exec, os.system, or subprocess calls, marking them with is_safe=False for optional exclusion.

⚖️ License & Terms

This dataset is free to use for any purpose (commercial, academic, or personal) without requiring a paid license. The only condition is mandatory attribution:

"PyMAX by Inserloft Research (Research.Inserloft.com)"

You must include this citation in your System Card, Model Card, or documentation where you reference public datasets used.

🏛️ Citation

If you use PyMAX in your research or publications, please cite:

@misc{pymax2026,
  title={PyMAX: Ultra Dataset for Python Learning},
  author={Inserloft Research},
  year={2026},
  month={August},
  url={https://huggingface.co/datasets/Inserloft/PyMAX},
}

👥 About Inserloft Research

Inserloft Research is an AI research organization focused on creating high-quality datasets and models for programming and software development. We specialize in synthetic data generation, code intelligence, and LLM fine-tuning.

🌐 Website: Research.Inserloft.com

📧 Contact: Inserloft@gmail.com

📜 Changelog

2026-08-24 (v1.0): Initial release (250M examples, 8GB).
2026-08-24 (v2.0): Major upgrade: 500M examples, AST validation, enriched metadata, ChatML format, test cases, safety flags, and optimized generation pipeline.

Downloads last month: 193

Total file size:

62.2 GB