ddvd233/hb27b_specgap_simple_retrieval_aicr_global_step_60

hb27b_specgap_simple_retrieval_aicr (global step 60)

Merged HF weights (bf16 safetensors) of the verl FSDP checkpoint hb27b_specgap_simple_retrieval_aicr/global_step_60.

• Base model: Qwen/Qwen3.6-27B
• Experiment: RRIMed paper baseline: identical RL recipe with the reward held fixed (fixed prompt), 27B, ARM 18
• Score: HealthBench Professional length-adjusted accuracy 0.370 at this step (run best 0.392 at step 75, rotated away; untrained 0.374)
• Note: Fixed-prompt control for RRIMed-27B.

This is the best checkpoint of the run whose weights survived checkpoint rotation; the run's best-validation step is stated above when it differs.

Research artifact trained on model-written tasks and evaluated on one benchmark family. Not for clinical use.