ddvd233/hb27b_specgap_ship_retrieval_websearch_global_step_200

hb27b_specgap_ship_retrieval_websearch (global step 200)

Merged HF weights (bf16 safetensors) of the verl FSDP checkpoint hb27b_specgap_ship_retrieval_websearch/global_step_200.

• Base model: Qwen/Qwen3.6-27B
• Experiment: RRIMed-27B (self-evolving rewards), main HealthBench Professional run, ARM 16
• Score: HealthBench Professional length-adjusted accuracy 0.485 at this step (run best 0.516 at step 120, whose weights were rotated away; untrained 0.374; fixed-prompt baseline 0.392)
• Note: Used for the HealthBench Hard transfer and MedXpertQA transfer rows.

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.