geopti/github-repos-8star-readmes

GitHub Repositories (≥8 stars) with READMEs

3,343,186 GitHub repositories — every public repository with 8 or more stars as of mid-July 2026 — each with its README text plus star count, description, primary language, creation date, and maintenance/fork metadata.

Collected June–July 2026 via the GitHub GraphQL API, then refreshed on 2026-07-16 to add repositories that crossed the 8-star threshold since the initial scan, and de-duplicated for renamed/transferred repositories.

Files

File	Size	Contents
repos_graphql.jsonl	~18 GB	The dataset: one JSON object per line, one repository per line
repos_index.csv	~308 MB	Lightweight rank index: rank, full_name, stars, language, has_readme, url (sorted by stars)
renamed_dupes.jsonl	0.7 MB	Audit report of the rename/transfer de-duplication pass
scripts/		The collection pipeline (see below)

Record schema (repos_graphql.jsonl)

Field	Description
full_name	owner/name (canonical nameWithOwner)
name, owner	Split of the above
stars	Stargazer count at fetch time
description	Repository "about" blurb
language	Primary language as reported by GitHub
url	https://github.com/<full_name>
created_at	ISO-8601 creation timestamp
default_branch	Default branch name
readme_content	Full README text (Markdown/plain), null if none found
has_readme	Whether a README was found (coverage: 93.7%)
is_archived	Owner archived the repository
pushed_at	Last push timestamp — the best "is it maintained" signal
is_fork, fork_parent	Fork flag and the owner/name it was forked from
forks	Fork count
disk_usage_kb	Repository size in KB
license	SPDX id (e.g. MIT, Apache-2.0) or null

Read it streaming — never json.load the whole file:

import json
with open("repos_graphql.jsonl") as f:
    for line in f:
        repo = json.loads(line)

How it was collected

GitHub's search API returns at most 1,000 results per query, so the scraper (scripts/scrape_streaming.py, reusing the query/fetch logic of downaload_github_readmes.py) enumerates repositories by star band and, within each band, by creation-date window, splitting windows until every query returns fewer than 1,000 results. Each query fetches repository metadata and the README blob in a single GraphQL call. Windows are journaled as they complete, so the multi-week run survived interruptions and resumed mid-band. refresh_8star.py later enumerated all current ≥8★ names, diffed them against the dataset, and full-fetched only the ~42k missing ones; find_renamed_dupes.py then removed 4,824 stale records for repositories that had been renamed or transferred (identified by identical created_at, confirmed via the API's rename-following repository() lookup). Verification against live GitHub counts showed every star band at 99–100% completeness.

Caveats

• Star counts are a snapshot at fetch time; a small number of records near the 8-star boundary have drifted above or below it since.
• ~6% of repositories have no README under any common spelling.
• README contents are the property of their respective repository owners and are subject to those repositories' own licenses; this dataset compiles publicly available data under the GitHub Terms of Service. Access is gated to keep usage accountable.