wckwan/WebShop-Olmo3-7B-Adaptive-Pivot-Fenced-eval

WebShop OLMo-3-7B-Instruct pivot FENCED: held-out evaluation

WARNING: global_step_140 is the odd row out. It was produced on the NEW action parser (verl-agent commit 451b4969, which reads the action after the final </think>). It is NOT comparable to global_step_20–global_step_120, which ran on the frozen OLD-parser surface 7173787ad6e171c6 (first <action> pair anywhere). config_fingerprint does not distinguish parser surfaces, so the rows look comparable but are not. The old parser corrupts fenced decisions increasingly with training step, so mixing the two surfaces would produce a step change at gs140 that is pure parser.

The new-parser series (gs140 onward) lives in wckwan/WebShop-Olmo3-7B-Adaptive-Pivot-Fenced-eval-newparser. Each row states its surface in result.json: eval_contrast_note on gs20-120, eval_surface_note on gs140.

Layout: global_step_N/{result.json, trajectories.jsonl, eval_seed0.log}, MANIFEST.sha256 (sha256sum -c compatible).