You can deploy this model and start earning money today!

mellee030/
MedQwen-3B-LoRA-r16

Name: MedQwen-3B-LoRA-r16
Availability: InStock

Qwen/Qwen2.5-3B-Instruct

adapter

apache-2.0

Views

No views yet

MedQwen-3B-LoRA-r16

LoRA fine-tuned Qwen2.5-3B-Instruct on 30K Chinese medical Q&A dialogue pairs with rank-16 adapters.

Evaluation Results

Metric	Base	Fine-tuned	Δ
ROUGE-1	0.0160	0.0227	+0.67%
ROUGE-2	0.0000	0.0014	+0.14%
ROUGE-L	0.0160	0.0193	+0.33%
BERTScore	0.6159	0.6751	+5.92%

BERTScore uses bert-base-chinese. ROUGE uses character-level tokenization for Chinese.

Rank Ablation (Qwen2.5-3B, lr=2e-4)

Rank	Trainable Params	BERTScore (ft)	Δ
r=8	~15M	0.6683	+5.24%
r=16	~30M	0.6751	+5.92%

r=16 outperforms r=8 by +0.68% BERTScore, showing that higher adapter expressiveness benefits Chinese medical Q&A fine-tuning when learning rate is held constant.

Model Comparison

Model	Rank	Fine-tuned BERTScore	Δ
MedQwen-1.5B	r=8	0.6620	+5.61%
MedQwen-3B	r=8	0.6683	+5.24%
MedQwen-3B	r=16	0.6751	+5.92%
MedQwen-7B	r=8	0.6670	+7.11%

Training Configuration

Parameter	Value
Base model	Qwen2.5-3B-Instruct
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training samples	30,284
Validation samples	412
Max sequence length	512
Batch size (effective)	8
Learning rate	2e-4
LR schedule	Cosine with warmup
Early stopping	patience=5, eval every 500 steps
Hardware	GCP L4 24GB
Precision	FP16

Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen2.5-3B-Instruct"
adapter_id    = "mellee030/MedQwen-3B-LoRA-r16"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id, dtype=torch.float16, device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

messages = [
    {"role": "system", "content": "你是一个专业的医疗问答助手，请根据用户的问题给出准确、简洁的医疗建议。"},
    {"role": "user",   "content": "糖尿病的早期症状有哪些？"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Dataset

30,284 training pairs and 412 validation pairs from a Chinese medical Q&A corpus, formatted as Qwen2.5 chat template (system / user / assistant turns).

Disclaimer

This model is for research purposes only and should not be used as a substitute for professional medical advice.