Benchmarks

Headline results, comparisons against baselines, and the interactive knowledge dashboards live on the paper page — this page is about running the reference benchmarks yourself.

Reference-benchmark data roots, data-prep scripts, task maps, and run presets live under benchmarks/ at the repo root — outside this docs site's build (docs/), so it's linked here rather than duplicated.

Start with the entry point:

• benchmarks/README.md — layout, run presets, where results land, and the licensing/attribution note for the third-party task corpora.
• benchmarks/docs/BENCHMARK_PREPARE.md — dataset preparation and task-map details for each benchmark.

For the CLI flags and concurrency knobs that apply regardless of task source, see Running larger runs.