Qwen2.5-7B-ReasonMed-cot-123k

Full fine-tune of Qwen/Qwen2.5-7B on the CoTMed variant of ReasonMed — chain-of-thought reasoning followed by the answer, without <think> tags. 123K samples (1/3 of full dataset) × 3 epochs.

Training code: https://github.com/Chen-Jie7/NLP_project

Training data

lingshu-medical-mllm/ReasonMed — CoTMed.json variant. Outputs are free-form CoT reasoning concluding with the answer.

Three-way format comparison

All three models trained with identical hyperparameters on 123K samples. Evaluated via loglikelihood MCQ scoring.

Variant	Output format	Total acc
reason	<think>CoT</think>Response	65.8
cot (this model)	CoT without tags	65.0
response	Direct answer only	63.8

Evaluation

Benchmark	Ours (123K)	Paper (370K)
MedQA	61.0	66.9
MedMCQA (val)	60.4	65.1
PubMedQA	75.9	82.0
MMLU-Anatomy	74.8	75.6
MMLU-Clinical-Knowledge	78.1	79.3
MMLU-College-Biology	81.9	79.2
MMLU-College-Medicine	68.8	73.4
MMLU-Medical-Genetics	84.0	85.0
MMLU-Professional-Medicine	79.0	80.9
Total	65.0	69.6

Training hyperparameters

• learning_rate: 1e-05
• effective batch size: 128 (8 GPU × 4 per-device × 4 accum)
• num_epochs: 3.0
• lr_scheduler: cosine, warmup_ratio 0.1
• precision: bf16
• deepspeed: ZeRO stage 2
• hardware: 8× H200, ~6.5h

Framework versions

• Transformers 4.57.6
• Pytorch 2.10.0+cu128
• LlamaFactory 0.9.5