GRPO LoRA adapter (r=32, α=64) trained directly on top of Qwen2.5-14B-Instruct (no continual pre-training), using the HealthBench-BR train split as reward. Reference for the ablation that quantifies the contribution of CPT.

Base model: Qwen/Qwen2.5-14B-Instruct
Type: LoRA adapter (PEFT)

Test-split accuracy

Benchmark	Accuracy
HealthBench-BR	64.2%
PCDT-QA	29.4%

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct", torch_dtype="auto", device_map="auto")
tok  = AutoTokenizer.from_pretrained("hugo/protocolos-clinicos-br-base_rl-1gen-14b")
model = PeftModel.from_pretrained(base, "hugo/protocolos-clinicos-br-base_rl-1gen-14b")

Intended use & limitations

Research model for studying domain adaptation of LLMs to Brazilian clinical guidelines. Not a certified medical device. Even at the best accuracy reported in the paper, residual errors may involve consequential details (dosages, contraindications). Use only under qualified professional supervision.

Citation

See the paper and code at the project repository:

Code & paper: https://github.com/hugoabonizio/clinical-protocols-br