Groundtruth Dynamic Benchmarking — Leaderboard
Groundtruth is a domain-specific LLM benchmark for geological reasoning, evaluating how well AI models reason over real geoscience source material (mineral exploration reports, government geological archives, and technical filings) rather than textbook recall. Candidate models answer fixed district question sets, and an independent LLM judge grades every answer 0–10 against a calibrated, source-anchored rubric.
Self-reported pointwise rubric scores (0–10 per question) against the groundtruth-dynamic-benchmarking geology rubrics, read live from public datasets. Nothing is hardcoded on this page.
Results are self-reported, not independently re-run. Each row is exactly what a submitter's own scoring run produced; nothing here is re-verified server-side. Use the "source" link on each row and the linked model/harness repo to check reproducibility yourself.
This is a pointwise leaderboard: each row is one model's (or harness's)
average score against one rubric, not a head-to-head comparison. Rows
whose question ids don't match the rubric named in their own
meta.json are flagged, not hidden.
By harness fixes the model (see
harness_track.json in the benchmark dataset) so only rows
calling that exact model are shown, ranked by harness rather than by
model — see the submissions dataset README for how to enter.