Full fine-tune of Qwen/Qwen2.5-7B on the ReasonMed dataset (<think>CoT</think>Response format), using a 123K-sample subset (1/3 of the full 370K dataset) for 3 epochs.

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

Training data

lingshu-medical-mllm/ReasonMed — ReasonMed.json variant. Each sample wraps chain-of-thought reasoning in <think>...</think> tags followed by the final answer.

Evaluation

Loglikelihood multiple-choice scoring across 9 medical QA benchmarks (eval_ll.py in the repo). Paper baseline is ReasonMed-7B trained on the full 370K samples.

Benchmark	Ours (123K)	Paper (370K)	Δ
MedQA	62.2	66.9	-4.7
MedMCQA (val)	60.9	65.1	-4.2
PubMedQA	77.5	82.0	-4.5
MMLU-Anatomy	75.6	75.6	0.0
MMLU-Clinical-Knowledge	78.5	79.3	-0.8
MMLU-College-Biology	81.9	79.2	+2.7
MMLU-College-Medicine	70.5	73.4	-2.9
MMLU-Medical-Genetics	84.0	85.0	-1.0
MMLU-Professional-Medicine	79.0	80.9	-1.9
Total	65.8	69.6	-3.8

The gap is consistent with training on 1/3 of the data.

Training hyperparameters

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

Framework versions

Transformers 4.57.6
Pytorch 2.10.0+cu128
Datasets 4.0.0
Tokenizers 0.22.1

Chens7/Qwen2.5-7B-ReasonMed-reason-123k

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

Training data

Evaluation

Training hyperparameters

Framework versions

Chens7/
Qwen2.5-7B-ReasonMed-reason-123k