Pairwise process reward models from MedCoast, fine-tuned from answerdotai/ModernBERT-base. They score partial reasoning chains for medical QA. The checkpoints are for inference only.
Inputs are capped at 2048 tokens.
| File | dsim | a | val pairwise acc |
|---|---|---|---|
medcoast_prm_dsim0.75_a0.0.pt |
0.75 | 0.0 | 0.5995 |
medcoast_prm_dsim0.75_a0.5.pt |
0.75 | 0.5 | 0.5757 |
medcoast_prm_dsim0.75_a1.0.pt |
0.75 | 1.0 | 0.5225 |
medcoast_prm_dsim0.8_a0.0.pt |
0.8 | 0.0 | 0.6035 |
medcoast_prm_dsim0.8_a0.5.pt |
0.8 | 0.5 | 0.5590 |
medcoast_prm_dsim0.8_a1.0.pt |
0.8 | 1.0 | 0.5143 |
medcoast_prm_dsim0.85_a0.0.pt |
0.85 | 0.0 | 0.6093 |
medcoast_prm_dsim0.85_a0.5.pt |
0.85 | 0.5 | 0.5411 |
medcoast_prm_dsim0.85_a1.0.pt |
0.85 | 1.0 | 0.5101 |
Each .pt file is a dict with model_state (the PRM state_dict, fp32) and config.
Use it with PRMScorer from the MedCoast codebase (model/prm_scorer.py):
from huggingface_hub import hf_hub_download
from model.prm_scorer import PRMScorer
ckpt = hf_hub_download("zhongys97/MedCoast", "medcoast_prm_dsim0.75_a0.0.pt")
scorer = PRMScorer(api_keys_json_path="api_keys.json",
base_model_name="answerdotai/ModernBERT-base",
ft_model_path=ckpt)
medcoast_prm_dsim0.75_a0.0.pt 596 MBmedcoast_prm_dsim0.75_a0.5.pt 596 MBmedcoast_prm_dsim0.75_a1.0.pt 596 MBmedcoast_prm_dsim0.85_a0.0.pt 596 MBmedcoast_prm_dsim0.85_a0.5.pt 596 MBmedcoast_prm_dsim0.85_a1.0.pt 596 MBmedcoast_prm_dsim0.8_a0.0.pt 596 MBmedcoast_prm_dsim0.8_a0.5.pt 596 MBmedcoast_prm_dsim0.8_a1.0.pt 596 MB