llmbench

LLM benchmark results from every listed leaderboard, side by side, at llmbench.site. Every number keeps its source link, and you choose which columns to compare.

The design and roadmap are in docs/llm-benchmark-aggregator-plan.md. Phases 0–3 are built and Phase 4 is under way: 30 benchmark sources, a model catalogue merged from whatmodel.sh, models.dev, Epoch AI's models database and Hugging Face, change history, checks on every update, a scheduled GitHub workflow that commits new data and deploys the site, the explorer with a chart mode, a page for every model, benchmark and source, model-name matching you can correct in one line, downloads, an in-browser SQL console, and a command that drafts a new source's entry.

The site

• Explorer (home): pick any columns from any source, or start from a preset such as Coding or Math. Filter by release date, size, maker, licence and who ran the evaluation. Sort by any column, click a score to see every score behind it and where each came from, and compare up to five models side by side, or switch to a chart of any column against release date, size or another column. The whole view is in the URL, so "Copy link" shares exactly what you see, and "Download CSV" exports it.
• Model pages (/models/<id>/): every score for one model, with its rank in each column.
• Benchmark pages (/benchmarks/<source>/<metric>/): the full leaderboard for one column, and where other sources have the same benchmark, every model's score from each of them side by side. The benchmarks page lists those benchmarks, each with its largest median gap between sources.
• Source pages (/sources/<id>/): status, licence, metrics, how each model name was matched, and recent changes.
• Downloads (/data/): every score as CSV, Parquet and SQLite, with models, benchmarks, names and sources.

How it works

1. You list sources in sources.yaml.
2. The update workflow runs every 6 hours. It fetches each source that is due, checks the result, records what changed, and commits it to data/.
3. The same run builds the site from data/ and deploys it to GitHub Pages.
4. If a source breaks, its last good data stays published and an issue titled Source problem: <id> is opened. It closes itself when the source recovers.

Adding a source

Most sources need no code. For a CSV, JSON, YAML, Parquet file or a Markdown table (such as a leaderboard in a README), let the tool draft the entry:

npm run add-source -- https://example.org/leaderboard.csv
npm run add-source -- github:owner/repo/README.md --table 0
npm run add-source -- https://example.org/leaderboard.csv --ai

It reads the file, lists its columns, and prints an entry to check and paste into sources.yaml. With --ai, Claude drafts the mapping instead (labels, units, rows to leave out, who ran the evaluations) from the columns and the first 25 rows, and the draft is run on every record before you see it; problems go back to Claude to fix. This sends that sample to the Claude API, so it needs ANTHROPIC_API_KEY or a profile from ant auth login; --model picks another model. Either way the entry comes with terms: blocked until you have read the source's terms. A source is blocked only when its terms forbid reuse or automated access; a missing or non-commercial licence is not a reason, and is labelled on the site and in every download instead. A finished entry looks like this:

  - id: my-board
    name: My Leaderboard
    homepage: https://example.org/leaderboard
    fetch: https://example.org/leaderboard.csv     # or github:owner/repo/path, or hf:owner/dataset
    model: Model                                   # the column with the model name
    license: MIT                                   # as the source states it, or "none stated"
    terms: allowed                                 # only after you have read the site's terms

Every numeric column becomes a score unless you list scores. Long tables (one row per model and metric) use metric and value. The comment at the top of sources.yaml lists every option, and the plan's section 3 explains them. Sources that need custom logic get a plugin in scripts/sources/ (see epoch.ts).

Check a new entry locally before committing:

npm run validate
npm run scrape -- --source my-board --dry-run   # shows what would be recorded, writes nothing

Run it locally

Needs Node 24 or later (it runs the TypeScript directly) and Python 3 for the preview server.

npm install
npm run scrape    # fetch due sources into data/ (--force to ignore "not due" and "unchanged")
npm run resolve   # match model names to the catalogue
npm run build     # build site/dist from data/
npm run preview   # serve site/dist at http://localhost:4321

npm test runs the tests and npm run typecheck type-checks.

The model catalogue

Each model's maker, release date, size, licence and context window come from catalogue feeds, merged in the order sources.yaml lists them: whatmodel.sh first, then models.dev and Epoch AI's models database, which add models released since, then Hugging Face, which only fills in parameter counts and licences for open-weights models. The first feed to describe a model wins; later feeds fill only what it left empty. A later feed's model is added only when nothing in the catalogue shares its core name, so a new feed never moves a name that already matched; the ones left out are listed in data/reports/catalog-left-out.csv (scripts/lib/catalog.ts has the rules). To take one that is a release of its own, add a line with just its id to curation/models.yaml; the feed's entry is used.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗