Datasets:

jblitzar
/

github-python-meta-elaborated

dataset_info:
  features:
    - name: repo_owner
      dtype: string
    - name: repo_name
      dtype: string
    - name: file_path
      dtype: string
    - name: file_url
      dtype: string
  splits:
    - name: train
      num_bytes: 24941794
      num_examples: 186066
  download_size: 8323253
  dataset_size: 24941794
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
annotations_creators:
  - author
license:
  - gpl-3.0
multilinguality:
  - monolingual
pretty_name: GitHub-Python
dataset_name: github-python
dataset_type: code
tags:
  - code
  - python
size_categories:
  - 100K<n⩽1M
task_categories:
  - text-generation

GitHub-Python — Licensed & Elaborated Variants

This repository ships two complementary Python-code corpora extracted from public GitHub:

Licensed Subset – strictly permissive-licensed files suitable for commercial redistribution / model training (main corpus used in our experiments).
Elaborated Collection – a broader crawl that additionally contains files under copyleft or unclear licenses (GPL/AGPL/LGPL, etc.). Useful for analysis or pre-training where license mixing is acceptable.

Both variants target code-completion / generation research.

Dataset at a glance

	Licensed Subset	Elaborated Collection
Files (.py)	53,017	186,066
Unique repositories	16,447	59,852
Repository owners	12,515	43,517
Compressed size	732 MB	2.4 GB *
Vocabulary (tokens)	443,431	443,431 †
License coverage	Permissive only	Mixed (perm. + copyleft)
Secrets redacted	✅	⚠️ not guaranteed
Time window	≥ 2015-01-01	≥ 2015-01-01

* estimated – elaborated corpus is distributed as raw file list, not a single text file.
† same tokenizer file is shared by both variants.

Numbers were obtained from the final redacted corpus and companion metadata.

Dataset structure

huggingface_dataset/
 ├─ mega_licensed_corpus_redacted.txt      # Licensed Subset – concatenated code
 ├─ python_files.txt                       # Licensed Subset – raw file URLs
 ├─ python_files_elaborated.txt            # Elaborated Collection – raw file URLs
 ├─ python_files_elaborated_metadata.csv   # Elaborated Collection metadata
 └─ custom_tokens_vocab.txt             # `<token>\t<id>` vocabulary file

File separator

Individual files are concatenated with the sentinel line:

# <FILESEP>

Anything following the sentinel until the next sentinel (or EOF) is the source code of one file.

Dataset variants

1. Licensed Subset (`mega_licensed_corpus_redacted.txt`)

• 53 K permissively-licensed files (MIT/BSD/Apache/ISC/Unlicense).
• All API keys & credentials removed.
• Ready for redistribution & commercial use (respect upstream NOTICE files).

2. Elaborated Collection (`python_files_elaborated.txt`)

• 186 K files from a much larger crawl.
• Contains GPL / LGPL / AGPL and other copyleft licenses.
• Shipped as URL list + metadata CSV; you must download the files yourself (datasets.load_dataset streaming, wget, etc.).
• No license filtering or secret-redaction performed – use with caution.

When first loading the dataset, decide which variant aligns with your use case (e.g. proprietary model training → Licensed Subset only).

Collection methodology

Repository discovery
- Queried GitHub REST API for projects with ≥ 10 stars
  (earlier iterations used 100+, later expanded for coverage).
- Only repositories with primary language Python and last commit ≥ 2015.
File filtering
- Retain files whose size ∈ [1 KB, 100 KB].
- Exclude common build/packaging scripts (setup.py, __init__.py, etc.).
License compliance
- Allowed: MIT, Apache-2.0, BSD-2/3-Clause, ISC, Unlicense.
- GPL, LGPL, AGPL and proprietary licenses were excluded.
Deduplication
- Unique file SHA hashes; duplicates skipped.
Formatting & cleaning
- Formatted with autopep8 to normalise whitespace.
- Custom script removed trailing whitespace & normalised newlines.
Secret redaction
- truffleHog + custom regex pass removed >150 active credentials.
- Redacted corpus stored as mega_licensed_corpus_redacted.txt.

Custom tokenisation

The accompanying custom_tokens_vocab.txt implements a Python-aware sub-token scheme:

Strip doc-strings & comments.
Split on:
- Camel-Case boundaries (Camel → Camel, Case)
- Underscores, spaces
- Indentation & newlines (preserved as <newline> token)
Rare tokens (frequency < 10) were dropped → 443 k vocabulary.

Example:

def helloWorld(value):
    return value + 1

tokenises to:

def hello world ( value ) <newline>     return value + 1 <newline>

Usage

from datasets import load_dataset

ds = load_dataset("jblitzar/github-python", split="train")

print(ds[0]["code"][:300])        # raw source code

If you prefer token level examples (small reasons: memory), map the tokenizer:

from tokenizers import Tokenizer
tok = Tokenizer.from_file("custom_tokens_vocab.txt")

def encode(ex):
    ex["input_ids"] = tok.encode(ex["code"]).ids
    return ex

ds = ds.map(encode, remove_columns=["code"])

Ethical considerations & limitations

Licenses respected – only permissive licenses included; retain NOTICE files when redistributing derivative works.
Secrets removed – automated & manual audits performed, yet users must not assume zero secrets; re-audit before public deployments.
Code quality – projects vary in style & correctness. Generated models may replicate bugs or vulnerable patterns.

Citation

If you use this dataset, please cite:

@misc{github-python-2024,
  author       = {JBlitzar},
  title        = {GitHub-Python: A Permissively Licensed Corpus of Python Code},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/jblitzar/github-python}},
  note         = {Version 1.0}
}

License

Dataset card and aggregation scripts: GPLv3.
Each code snippet remains under its original repository license (MIT, Apache-2.0, BSD, ISC, etc.). Users must comply with upstream notices when redistributing code or derivatives.