license: apache-2.0
base_model: FreedomIntelligence/HuatuoGPT-o1-8B
datasets:
- BrainHealthAI/MedCortex-v1
language:
- en
tags:
- medical
- clinical-reasoning
- chain-of-thought
- full-fine-tune
pipeline_tag: text-generation
MedCortex-Huatuo-8B-e1
Full fine-tune of FreedomIntelligence/HuatuoGPT-o1-8B on the reasoning subset of BrainHealthAI/MedCortex-v1 (English, verified chain-of-thought, Apache-2.0).
This reproduces the MedReason recipe (arXiv:2504.00993) with the only variable being the training
data: same base, same official SFT.py recipe, same official evaluation harness. It is an
internal research checkpoint for a controlled post-training evaluation.
Training
- Base: FreedomIntelligence/HuatuoGPT-o1-8B (itself Llama-3.1-8B + medical RL)
- Data: MedCortex-v1
task_type=reasoning, English (42086 examples) - Recipe: full FT (no LoRA), 1 epoch, LR 5e-6, cosine, warmup 0.05, global batch ~128, bf16, DeepSpeed ZeRO-3, seq_len 4096, seed 2002
- Response format:
## Thinking\n\n<reasoning>\n\n## Final Answer\n\n<answer>
Prompt format
Use the Llama-3.1 chat template; ask a clinical/exam question and the model reasons under
## Thinking then answers under ## Final Answer.
Evaluation
Evaluated with the official MedReason harness (SGLang + eval.py + scorer.py, strict prompt),
so scores are drop-in comparable to MedReason paper Table 4.
Intended use & limitations
Decision support / research only — not a medical device; every output requires a licensed professional. 1-epoch checkpoint (paper uses 3): an early read on data quality, not a final model.
Citation
@misc{wu2025medreason, title={MedReason: Eliciting Factual Medical Reasoning Steps in LLMs via Knowledge Graphs},
author={Juncheng Wu and others}, year={2025}, eprint={2504.00993}, archivePrefix={arXiv}}