A live map of all official benchmark leaderboards on Hugging Face: which fields they cover, how many models take part, and who submits. Unofficial; read straight from the public Hugging Face API each time you open the page.
How many official benchmarks each field has, and how many leaderboard entries they have received.
Leaderboard entries on every official benchmark, busiest first, colored by field.
Leaderboard entries per organization (top 20).
All leaderboard entries by country of the publishing organization.
Per organization.
Each official leaderboard with entries counts once per organization at rank 1.
Country of the organization that published the leading model.
Number of #1 spots per organization in each field. Columns: organizations with at least one #1.
Relative gap between #1 and #2 on each leaderboard, in percent of the #2 score (direction-aware: on lower-is-better boards a lower #1 score counts as ahead). Boards with a single entry are left out.
What this is. Hugging Face marks some datasets as official benchmarks and shows a leaderboard on each, built from the .eval_results files that model repositories publish. This page lists those benchmarks and counts who leads them.
Data. /api/datasets?filter=benchmark:official for the list and /api/datasets/<id>/leaderboard for each board. Nothing is cached or edited; everything is computed in your browser on load.
Read with care. Most entries are self-reported by model authors, and evaluation settings differ between submissions (sampling, majority voting, thinking budget). A verified tag appears only where Hugging Face marks an entry as verified. Organization = the namespace of the model repository; a fine-tune published under another namespace counts for that namespace. Ties at rank 1 count for every tied organization.
Fields and countries. These two labels are added by hand (field per benchmark, country per organization, based on where the organization is headquartered). Organizations without a clear public headquarters are shown as Unknown. The full mapping is in maps.js; corrections are welcome in the Community tab.
Built by QuantID. Not affiliated with or endorsed by Hugging Face.