Result snapshots and supporting data for the SkillsBench leaderboard.
The latest experimental results are the HuggingFace dataset
benchflow/skillsbench-leaderboard,
main branch + PR #11
(v0.1 + v1.1). It covers 18 model–harness configurations × 87 tasks × 2 conditions
(with / without curated Skills) × 3 trials, run at each model's maximum reasoning effort and
scored pass/fail. That dataset is the source of truth for both trajectories and results.
Build-time data consumed by skillsbench.ai (fetched from this repo's main):
results-registry.json— per(task, model–harness, condition)aggregate (score, pass/perfect counts) backing each task page's Results board. Regenerated from the HuggingFace snapshot above.
The website no longer hosts agent trajectory files; trajectories are browsed and downloaded on the HuggingFace dataset.
The contributor-named directories (yimin/, xiangyi-*/, shenghan/, wenbo/,
minimax-m2.1+claude-code/) are an archive of earlier-generation runs. Current
trajectories live on the HuggingFace dataset, not in this repo.