Reproducing ClawBench V2 scores: corpus and evaluation protocol

#1
by Reacherx - opened
TIGER-Lab org

Maintainer note for users of this leaderboard Space. ClawBench V2 contains 130 live-website tasks; the public repository currently exposes V1+V2 task definitions (283 tasks total), while leaderboard rows identify the corpus and harness used. Scores are reported with both request-interception evidence and the agentic judge; please cite the exact corpus, harness, model, judge mode, and commit when adding a run.

Reproduction resources:

This note is informational and does not alter leaderboard results. New results should follow the repository submission instructions and include enough metadata for an independent rerun.

Sign up or log in to comment