OPD-LLM v11

LoRA adapters on Qwen/Qwen2.5-1.5B-Instruct (SFT on MedMCQA/PubMedQA, then GRPO-style RL on MedQA-USMLE). The root of this repo is the RL adapter. adapter_sft/ is the SFT-only adapter, adapter_vl/ is a Qwen2-VL-2B LoRA trained on VQA-RAD / PathVQA.

Results (MedQA-USMLE test, n=200, greedy)

Stage	Accuracy	95% CI	Format fail
Pretrained	0.395	[0.33, 0.464]	0.06
After SFT	0.375	[0.311, 0.444]	0.08
After RL	0.455	[0.387, 0.524]	0.0
After RL + notes	0.435	[0.368, 0.504]	0.01

Vision: {'vqa_rad_yes_no_acc_on_100': 0.64, 'train_steps': 87}

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, "TidasSilva/OPD")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")

Limitations

Research use only. Not clinically validated, not approved for patient care. Exam-style MCQ accuracy only; differences under ~7 points at n=200 are within noise.