| --- |
| license: cc-by-nc-4.0 |
| pretty_name: GitHub Stars in Context |
| language: |
| - en |
| tags: |
| - github |
| - mining-software-repositories |
| - stars |
| - repositories |
| size_categories: |
| - 10M<n<100M |
| configs: |
| - config_name: repositories |
| data_files: repositories.parquet |
| default: true |
| - config_name: users |
| data_files: users.parquet |
| - config_name: portfolios |
| data_files: portfolios.parquet |
| --- |
| |
| # GitHub Stars in Context |
|
|
| A dataset of repository popularity across GitHub accounts. It links |
| **3,671,497 GitHub user accounts** to **53,990,808 rows of public, non-fork |
| repositories** they own, with each repository's star count. A row is one |
| repository under one account. |
|
|
| Provenance, checks and known limits are documented in |
| [PROVENANCE.md](https://github.com/NaiveNeuron/github-stars-in-context/blob/main/PROVENANCE.md). |
|
|
| Most GitHub datasets describe repositories one by one. This one keeps the |
| ownership context: every stored repository of an account, including those with |
| zero stars. That makes it possible to see what a star-threshold filter leaves |
| out, to compare concentrated and spread-out portfolios, and to group |
| repositories by owner when building samples or benchmark splits. |
|
|
| Processing code, reuse examples and reproduction instructions: |
| <https://github.com/NaiveNeuron/github-stars-in-context> |
|
|
| ## Files |
|
|
| | File | Rows | Contents | |
| | --- | ---: | --- | |
| | `users.parquet` | 3,671,497 | Account IDs, login, retrieval status, profile-availability flag, follower count | |
| | `repositories.parquet` | 53,990,808 | Owner key, GitHub repository ID, name, creation time, stars | |
| | `portfolios.parquet` | 3,671,497 | Per-account repository count, total and maximum stars, h-index, concentration | |
| | `data-dictionary.md` / `.json` | | Field meanings, types and NULL rules | |
| | `release.json` | | Version, licences, counts, collection dates and file checksums | |
| | `crosswalk.sql` | | View mapping local keys to GitHub IDs, names and URLs | |
| | `PARQUET-SHA256SUMS` | | Checksums of the files above | |
|
|
| `repositories.user_id` joins `users.id`. Both are local keys. `users.github_id` |
| and `repositories.repository_id` are GitHub's numeric IDs and join to other |
| resources such as GH Archive. |
|
|
| ## Quick start |
|
|
| Download the files once (1.57 GiB). This needs the Hugging Face CLI, or use |
| the `curl` commands in the code repository's README: |
|
|
| ```sh |
| hf download NaiveNeuron/github-stars-in-context --repo-type dataset --revision v1.1.0 --local-dir data |
| (cd data && shasum -a 256 -c PARQUET-SHA256SUMS) |
| ``` |
|
|
| Then query the local files, for example with DuckDB (`pip install duckdb`): |
|
|
| ```python |
| import duckdb |
| print(duckdb.sql(""" |
| SELECT count(*) FILTER (WHERE r.stars >= 10) AS kept, count(*) AS total |
| FROM 'data/repositories.parquet' r |
| JOIN 'data/users.parquet' u ON r.user_id = u.id |
| WHERE u.retrieval_status = 'complete' |
| """).fetchall()) |
| # [(1560798, 53987867)] |
| ``` |
|
|
| DuckDB can also read the files remotely through `hf://` paths, but a query |
| over the large table sends several thousand requests and can exceed Hugging |
| Face's limit for anonymous access. Download the files, or log in first. |
|
|
| ## How it was collected |
|
|
| - **Accounts** were found on 8–9 December 2025 through GitHub GraphQL user |
| search, split by follower count (1–9 up to 10,000+) and registration date |
| to stay under the 1,000-result search limit. Organizations are excluded. |
| Accounts with no followers were not searched |
| - **Repositories** were fetched per account from about December 2025 to |
| early January 2026, by the collector author's estimate. The query selects |
| owned, public, non-fork repositories, including archived ones, and follows |
| pagination. Per-request times were not stored |
| - **Follower counts** were fetched from 28 January to 2 February 2026 |
|
|
| ## What to know before using it |
|
|
| - **Not a random sample of GitHub.** Discovery searched for accounts with at |
| least one follower. Follower counts fetched later can be zero (1,719 |
| accounts). Search cells that still exceeded 1,000 results on a single |
| registration day were skipped, and a failed search request could drop a |
| cell. The number of missed accounts is unknown |
| - **Retrieval status.** `retrieval_status` is `complete` for 3,615,479 |
| accounts (372,821 of them own no stored repository), `failed` for 56,012 |
| and `unknown` for 6. The classes are derived from the collector's logic. |
| They are not an independent check against GitHub |
| - **Missing follower counts.** `profile_available` is false for 1,151,092 |
| accounts whose profile fields are missing or ambiguous. Their |
| `followers_count` is NULL. The rule masks NULL biographies and profiles |
| with only default values, which can include genuinely empty profiles. In a |
| sample of 300 such accounts of the complete class checked on 10 November |
| 2026, 298 still existed and 296 had followers. This supports treating the |
| stored zeros as missing, but it does not show their values in January 2026. |
| The raw stored value is kept in `followers_count_raw` and must not be read |
| as zero followers. Follower analyses cover only the accounts with available |
| profiles, measured after the repositories |
| - **Historical observations.** Star counts are as of retrieval. Per-request times were |
| not stored, so this is not a star history |
| - **Repository IDs are not unique keys.** Twenty repository IDs appear under |
| more than one account |
| - **Stars are not a measure of people.** Do not use this dataset to rank or |
| evaluate individuals |
|
|
| ## Personal data |
|
|
| The dataset contains only fields that were public on GitHub: account logins |
| and numeric IDs, repository names, creation times and star counts. Logins and |
| numeric IDs can identify people. Numeric IDs are needed to join the tables to |
| other datasets, and logins allow lookup by name. Biographies, company, |
| location and external links were collected but are **not** included. |
|
|
| To have an account removed, write to marek@suppa.sk or open an issue in the |
| code repository. Please do not post personal documents in a public issue. We |
| remove the account's rows from all three tables in a new release. Copies that |
| others have already downloaded cannot be recalled. |
|
|
| ## Licence |
|
|
| The data are released under |
| [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). The licence |
| covers the rights the authors hold in the compilation. It does not grant |
| rights held by GitHub users or other third parties. The code is MIT-licensed. |
|
|
| ## Versions |
|
|
| - **1.1.0** (2026-11-10): documentation only. The data dictionary is rewritten |
| in plain language, and `release.json` replaces `analysis-manifest.json`. |
| The three Parquet files are byte-identical to 1.0.0 |
| - **1.0.0** (2026-11-10): first public release |
|
|
| Each version is a Git tag of this repository (`v1.1.0`, `v1.0.0`). The code |
| repository has matching tags. |
|
|
| ## Citation |
|
|
| A paper describing the dataset is under review. Until it is published, |
| please cite this dataset page: |
|
|
| ```bibtex |
| @misc{pavlov2026githubstars, |
| author = {Pavlov, Mykhailo and {\v{S}}uppa, Marek}, |
| title = {{GitHub} Stars in Context: A Dataset of Repository Popularity Across Accounts}, |
| year = {2026}, |
| version = {1.1.0}, |
| howpublished = {\url{https://huggingface.co/datasets/NaiveNeuron/github-stars-in-context}} |
| } |
| ``` |
|
|
