PRBench · Legal
What it measures and how
Fix either all 500 Legal tasks or the 250 Hard tasks, with consistent search tools, reasoning tiers and judge versions. Use the task set’s aggregate rubric score. Do not double-count the full and hard sets or describe professional Q&A as complete file delivery.
How this evidence is used
Under observation: could add legal judgment evidence. Task sets, tool conditions and score batches need verification. Only the source is listed; no automated collection or ranking yet.
Limitations and attribution
Model judging introduces preferences. Legal and financial specialties from one organization share influence limits, rather than counting as two independent evidence sources. It does not directly measure Word, Excel or slide creation.
Data license: Official tasks and evaluation code are public. Their licenses do not automatically cover live leaderboard results. Stable exports and display terms for the latest results remain to be confirmed.
Results are published by Scale AI. Raw scores and the e/acc consensus index use different scales and cannot be added directly.