gpt56_sft_qwen36_27b (global step 28)
Merged HF weights (bf16 safetensors) of the verl FSDP checkpoint gpt56_sft_qwen36_27b/global_step_28.
- Base model: Qwen/Qwen3.6-27B
- Experiment: Reference baseline: Qwen3.6-27B supervised fine-tuned on GPT-5.6 traces for HealthBench Professional (distillation from a stronger teacher)
- Score: HealthBench Professional length-adjusted accuracy 0.419 under the official no-tools protocol (untrained 0.381)
- Note: The 9B counterpart is ddvd233/gpt56_sft_qwen35_9b_global_step_28.
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.