README.md

metadata

title: Grounding Leaderboard
emoji: 📊
colorFrom: blue
colorTo: green
sdk: static
app_file: index.html
pinned: false
license: cc-by-sa-4.0

Grounding Leaderboard

A static leaderboard for the grounding-benchmark: which NLI cross-encoders best tell whether a claim is actually supported by the document it cites, ranked by ROC-AUC (headline = number grounding), with inference speed on a standard NVIDIA RTX A6000.

• index.html — the page (fetches leaderboard.json).
• leaderboard.json — aggregated results (array of score.py outputs). Regenerated by scripts/score_leaderboard.sh on the standard RTX A6000.
• results/ — individual submission JSONs (one per model×config), for provenance and PRs.

Submit a model: score it with the reference scorer and open a PR adding your results/*.json. See SUBMISSION.md.