MESSIER v1.0.2
#3
by StefanKrsteski - opened
MESSIER v1.0.2
This release corrects the treatment of non-scorable attempts in Harbor evaluations.
- Corpus. 904,051 records from 31 benchmarks, 734 agents, 11,999 tasks, and 72,000 verifiers.
- Data. The release contains task definitions, results, verifier definitions, and 119,482 available trajectories.
- Harbor attempts. Attempts without verifier rewards are retained with
result=Noneinstead of being omitted. - Failure handling. Provider, environment, and evaluator failures are marked as non-scorable when they prevent evaluation, while agent timeouts and complete graded outputs remain scored.
- Analysis. The HarveyAI-Lab counterfactual analysis and derived corpus statistics are updated accordingly.
Published after the MESSIER v1.0.2 release PR was merged.
StefanKrsteski changed pull request status to merged