2026-10-21T17:02:05.466364
RL v26.00, steps 1–150. Raw task rewards from final step-156 trainer_state.json, with trailing 10-step means. Both BioReason panels share the single logged bioreason_task_reward, with separate coding and non-SNV evaluation scores. Checkpoint dots are latest successful primary scores at steps 30,60,90,120,150. VEPQA groundedness penalty excluded.
image/svg+xml
Matplotlib v3.10.8, https://matplotlib.org/