Hugging Face marks certain datasets as official benchmarks. Each one carries a leaderboard on its dataset page, built automatically from the .eval_results files that model repositories publish. GPQA Diamond, MMLU-Pro, Humanity's Last Exam, SWE-bench Verified, AIME 2026, HMMT February 2026, Terminal-Bench, MMMU-Pro, LEXam, the Open ASR Leaderboard, and more.
The problem: these leaderboards live on dozens of separate dataset pages. There is no single place that answers a simple question.
Who actually holds #1, across all of them?
So we built one.
What the index shows
It is a live, read-only view that recomputes from the Hugging Face API on every visit:
- #1 by organization — how many official leaderboards each organization currently leads, with a share chart.
- #1 by country — first-place spots grouped by the country of the publishing organization.
- Who leads which field — a heatmap of #1 spots across math, science and knowledge, coding, agents and terminal, documents and OCR, vision and video, speech, law and finance, retrieval and robotics.
- How far ahead #1 is — the relative gap between #1 and #2 on every board, direction-aware so that lower-is-better metrics like word error rate are handled correctly.
- Every benchmark — the #1 and #2 model and score, and the number of entries, for each board.
How it is built
Nothing is hand-entered. The page reads the list of datasets that Hugging Face tags as official benchmarks, pulls each board's ranked entries from the public API, and aggregates. Because a new #1 only appears after Hugging Face indexes the eval results, the index reflects the public record with the same lag as the source leaderboards.
A few design notes for anyone building something similar:
- Ties at rank 1 are counted for every tied organization.
- The #1-to-#2 gap is measured relative to the #2 score and is direction-aware.
- Country is derived from the publishing organization, not the model name.
Why it matters
Benchmark claims are easy to assert and hard to check. A single page that reads straight from the eval results turns "we lead on X" into something a reader can verify in one click, including how large the margin actually is. For teams, it is a fast way to see where the frontier sits in your field and who is defending it.
As of early October 2026 the index tracked around 49 official benchmarks with roughly a thousand leaderboard entries in total, and the single organization holding the most first-place spots was FINAL-Bench with eleven, spanning science, math, multimodal, law, document extraction, structured output and decision-making.
Open it, sort by field or country, and see for yourself.
Live index: https://huggingface.co/spaces/quantid/huggingface-official-benchmark-leaderboards
Top comments (0)