README.md

metadata

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - code
pretty_name: GoodDocs-v0
size_categories:
  - 100K<n<1M

GoodDocs-v0: High-quality code documentation texts

GoodDocs-v0 is a text dataset scraped from high-quality documentation sources in the open-source ecosystem, in particular the top 1000 GitHub repositories by stars. It is designed to serve as a foundation for building reasoning systems grounded in software documentation, enabling tasks such as:

• Code and API understanding
• Documentation question answering and retrieval
• Planning and tool-use grounded in docs
• Long-context reasoning over multi-file documentation

What's in this repository

• cleaned_texts_on_metadata_only.parquet — per-file Markdown documents and metadata extracted from documentation trees.
• awesome-repos.parquet — structured links extracted from Awesome lists-of-lists (name, link, description, source_repo, optional stars).
• data_collection_utils/ — utilities to regenerate the dataset:
  • scrape_gh_docs.py — main scraper/collector for documentation from GitHub repositories.
  • scrape_gh_docs_config.yaml — reproducible configuration (inputs, outputs, filters, strategies).
  • github_links.txt — the seed list of GitHub repositories (e.g., top repositories by stars).
  • awesome_final_repos.py — extractor for non-"awesome" repositories referenced by Awesome lists.
  • awesome_scrap_config.yaml — configuration for awesome_final_repos.py (root, depth, output, cache, workers, optional fetch_stars).
  • top_1000_repos.py — helper to refresh the top‑repositories list via the public site referenced in the code.

Schema

cleaned_texts_on_metadata_only.parquet — one row per Markdown file (see md_rows assembly in main()):

• owner, repo, repo_dir
• file_rel_repo — path relative to the saved repo root
• file_rel_outdir — path relative to outdir
• size — file size in bytes
• mtime — file modification time (epoch seconds)
• lang — language prediction field (via langid.py when language filtering is enabled)
• content — raw Markdown text

Quickstart

Load the dataset with pandas:

import pandas as pd
df = pd.read_parquet("cleaned_texts_on_metadata_only.parquet")
print(len(df), "rows")
print(df.columns.tolist())

Typical uses:

• Retrieval corpora for doc QA and RAG pipelines
• Supervision for instruction tuning grounded in docs
• Long-context model evaluation with real project documentation

Reproducing the dataset

The scraper is configurable and designed to be reproducible via data_collection_utils/scrape_gh_docs_config.yaml.

1. Prerequisites

   • System tools: git
   • Python 3.11+ packages: pandas, pyarrow, requests, tqdm, PyYAML, langid
   • For refreshing top repositories (optional): playwright (and playwright install for a browser)
   • A GitHub API token in the environment (GITHUB_TOKEN) or a file referenced by the config (token_file)

2. Inputs

   • data_collection_utils/github_links.txt — list of repositories to process (either owner/repo or full URLs)
   • You can refresh this list with data_collection_utils/top_1000_repos.py if desired.

3. Run

python3 data_collection_utils/scrape_gh_docs.py
# or to rebuild Parquet(s) from existing downloads without any network calls:
python3 data_collection_utils/scrape_gh_docs.py --no-fetch

Configuration (YAML-driven; see data_collection_utils/scrape_gh_docs_config.yaml):

• input — path to a file containing one repo per line (owner/repo or full URL)
• outdir, md_failed, texts_parquet
• workers, dry_run, quiet, no_fetch
• token_file — GitHub token location (or set GITHUB_TOKEN env var)
• prefer_sparse, prefer_zip, only_md, min_repo_age_years
• lang_filter, min_text_chars — control language gating in cleaned_texts_on_metadata_only.parquet

Output is written to <outdir>/cleaned_texts_on_metadata_only.parquet.

Awesome list extraction

data_collection_utils/awesome_final_repos.py crawls the Awesome list-of-lists and extracts final repositories (those whose repo names do not include "awesome"). For each bullet entry like:

* [Fuse](https://github.com/owner/repo) - Mobile development tools.

It records:

• name: the markdown link text (e.g., Fuse).
• link: canonical GitHub repository URL (e.g., https://github.com/owner/repo).
• description: text after the - dash, or the rest of the line (with the link and bullet removed) if no dash.
• stars (optional): repository stargazers count when enabled.

Configuration is YAML-first via data_collection_utils/awesome_scrap_config.yaml:

• root: root Awesome repository URL, e.g., https://github.com/sindresorhus/awesome.
• depth: recursion depth for nested Awesome lists (0 = only root).
• output_dir: directory for awesome-repos.parquet.
• cache_dir: directory for README fetch caches.
• workers: concurrency for network requests.
• fetch_stars: when true, also fetch stargazers for each parsed repo (makes extra API calls) and include a stars column.

Run:

python3 data_collection_utils/awesome_final_repos.py
# or adjust via YAML first, then run without flags

Schema of awesome-repos.parquet:

• name — link text from the Awesome entry.
• link — canonical GitHub URL (https://github.com/owner/repo).
• description — description text without the leading - and without repeating the name.
• source_repo — the Awesome list repository where the entry was found, formatted as owner/repo.
• stars — integer, optional; only present when fetch_stars: true.

Language filtering

Language detection is performed with langid.py (see imports in data_collection_utils/scrape_gh_docs.py). The default configuration keeps English-only files (lang_filter: en). There is no probability/confidence threshold; we gate by the predicted language label and a minimum text length (min_text_chars).

Licensing

• Code and dataset scaffolding in this repository are under the MIT license (see frontmatter).
• The original documentation content belongs to the respective upstream projects and remains governed by their licenses. Please consult each repository’s license before redistribution or commercial use.

Acknowledgements

This dataset draws from the open-source community’s documentation efforts. The seed list targets highly-starred repositories to bias toward quality, breadth, and maturity.

Note to self: size distribution: 20th percentile - 363 symbols, 50p - 701, 95p - 17392