---

license: mit task_categories: - text-generation - text-classification - question-answering language: - en tags: - code size_categories: - 10K<n<100K

PyHub: Vetted Python Code from Popular GitHub Repositories

A large-scale dataset of Python source code, test files, and documentation scraped from high-quality GitHub repositories. Designed for training code understanding and generation models, particularly for software engineering benchmarks like SWE-bench.

Dataset Statistics

• Total files: 1963
• Repositories: 25 (minimum 50 stars)
• File types: Python source, test files, READMEs
• Time period: Repositories created before January 1, 2020
• Size limit: Maximum 100 MB per repository
• License: MIT

Dataset Structure

Each row in dataset.csv represents a single file with the following columns:

Column	Description
repo_name	Repository name (e.g., "requests")
repo_full_name	Full repository name (e.g., "psf/requests")
owner	Repository owner (e.g., "psf")
stars	Star count
license	SPDX license identifier
repo_description	Repository description
filepath	Relative path within repository
file_type	"python", "test", or "readme"
language	"Python", "Markdown", "reStructuredText", or ""
content	Full file text
size_bytes	File size in bytes
num_lines	Number of lines

File Types

• Python source (.py): Production code files
• Test files (*test*.py): Unit tests and test suites
• README files (README.*): Documentation in Markdown, reStructuredText, or plain text

Collection Methodology

The dataset was collected using a custom GitHub scraper with the following process:

1. Repository selection: GitHub API search for repositories with x >= 50 stars, created before 2020-01-01, non-fork
2. Cloning: Shallow git clone (--depth 1) with 100 MB size filter to exclude large monorepos
3. File collection: Recursive walk through cloned repositories, excluding hidden directories (files starting with .)
4. File type filtering: Only Python source files (.py), test files (*test*.py), and README files (README.*) were collected
5. Content extraction: UTF-8 encoding with error handling for robust text extraction
6. Parallel processing: 3 concurrent workers for efficient processing
7. CSV generation: All file data consolidated into a single CSV with repository metadata embedded in each row

Quality Filters

• Star threshold: Minimum 50 stars (indicates community vetting)
• Size limit: 100 MB to exclude monorepos and binary-heavy projects
• File type filtering: Only Python, test, and documentation files
• Hidden files excluded: Files/directories starting with . ignored
• Encoding handling: UTF-8 with error fallback

Intended Use Cases

• Code completion: Training autocompletion models on real-world Python patterns
• Bug detection: Learning from production codebases with established testing practices
• Test generation: Understanding test-code relationships from included test files
• Documentation generation: Learning code-documentation correlations from READMEs
• SWE-bench training: Base dataset for software engineering benchmark preparation
• Code understanding: Repository structure and dependency learning

Limitations

• Temporal bias: Pre-2020 code, missing modern Python features (type hints, match statements, structural pattern matching)
• Popularity bias: High-star repos only, may not represent niche or edge-case patterns
• Size limitation: 100 MB cap excludes large enterprise monorepos
• Language bias: Primarily English documentation and comments
• Static only: No execution data, test results, or runtime behavior

Recommended Supplements

For comprehensive model training, consider supplementing with:

• Post-2020 repositories for modern Python patterns
• Smaller repositories for edge-case and niche patterns
• Synthetic examples for specific bug types
• Negative examples (buggy code) for robustness

License

This dataset is licensed under the MIT License. See the LICENSE file for details.

---

Note: This dataset was created using a custom GitHub scraper tool.