idealab-cs2/reappraisal-4b-grpo-committee
124
drop links to deleted near-duplicate HF repos
model card: exact cluster CIs (0.872 / CI-low 0.781; R1 h2h 0.673)
Trim Evaluation to card-appropriate length; keep honest facts, cut the stats lecture
Correct R1 overclaims: retract iid CI-low vs R1 reference-rate; state honest vignette-cluster CI-low and OOD parity/loss vs R1 (R1-OOD program closed negative)
Trim caveats to card-appropriate length
Card honesty pass: qualify judge calibration regime, note CI independence assumption + specialist framing, mark OOD gap as not significant
Match house writing style: remove non-native em dashes / flourishes
Update card: n=720 confirmed (0.872 CI-low 0.846) + direct-R1 head-to-head 0.671 + OOD 0.58 vs 0.52
Upload folder using huggingface_hub
initial commit
