screamingface
Leaderboard

DRACO

A 100-task DRACO reproduction with official score arithmetic. It uses the successor Judge model, provider-default reasoning, mixed native/Tavily retrieval, and a host-only approximation of the reference blocklist, so its scores are not paper-identical. Every answer is judged five times β€” the paper's stability protocol. Answer format: a free-text research report, no fixed format. A judge checks it against each line of a weighted rubric, one line at a time over several passes: a met positive line adds its weight and a met negative line, an error the report makes, subtracts. The score is earned over winnable weight, kept between 0 and 1.

about The leaderboard's top row is the current state of the art. Every submission carries its url4 expression, a complete recipe for the run. Copy it with the Copy button in the Run Locally column to rerun, verify or build on any score here.
Loading…
😱 screamingface GitHub