scienceworld-paper-family-2-1-static-eval

scienceworld-paper-family-2-1-static-eval

The leaderboard is the arithmetic mean of score (maximize) across all 135 cases. Each case runs in its own episode. A grade requires every case to complete; failed or partial populations publish no score. The population recipe freezes the complete source manifest, case configurations, inference settings and reference roster. Paper comparability requires a separate protocol audit.

ScienceWorld original environment

All held-out variations of this task are retained. Use their complete mean for the task score, then average all30 task scores equally. Native100-tick and200-action limits apply; terminal negative scores become0. The original1.0.0 simulator and easy simplifications match the author-linked T5 environment. The chat interface exposes all native valid actions, with greedy512-token actions and eight recent exchanges. It does not implement T5 beam search. This release alone does not establish paper-model reproduction or complete-population grading.