Qwen2.5-14B-Instruct continually pre-trained on synthetic data from a single generator (GPT-4.1-mini), mixed 1:1 with FineWeb (English) as replay corpus to mitigate catastrophic forgetting.

Base model: Qwen/Qwen2.5-14B-Instruct
Type: Full fine-tuned weights

Test-split accuracy

Benchmark	Accuracy
HealthBench-BR	63.8%
PCDT-QA	65.6%

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("hugo/protocolos-clinicos-br-cpt_replay-1gen-14b", torch_dtype="auto", device_map="auto")
tok   = AutoTokenizer.from_pretrained("hugo/protocolos-clinicos-br-cpt_replay-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