ggbaobao/medc_llm_based_on_qwen2.5

Model Details

This model has been LoRA‑fine‑tuned on Qwen2.5‑7B‑Instruct. In the future, reinforcement learning training may be carried out based on this model, such as DPRO algorithm, etc.

Base Model Sources [optional]

https://huggingface.co/Qwen/Qwen2.5-7B-Instruct

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "ggbaobao/medc_llm_based_on_qwen2.5"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16 
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "猩红热多在发热后多久出现皮疹，请从以下选项中选择：12小时之内, 12～48小时, 60～72小时, 84～96小时, 大于96小时"
messages = [
    {"role": "system", "content": "You are Qwen, You are a helpful assistant."},    
    {"role": "user", "content": prompt},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
    do_sample=True
)  
generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(response)

Training Details

lora_config = LoraConfig(
    r=16,              
    lora_alpha=32,    
    target_modules=["q_proj", "v_proj"],  
    lora_dropout=0.1 
)

training_args = TrainingArguments(
    output_dir="./results_final1",
    learning_rate=7e-5,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=1,  # 梯度累积
    num_train_epochs=2,
    evaluation_strategy="steps",
    # evaluate_steps=1,
    save_strategy="steps",
    save_steps=10,
    logging_steps=10,
    logging_dir="./logs1",
    bf16=True,  # 混合精度训练

Training Data

The training data comes from https://github.com/SupritYoung/Zhongjing If you want to know more details about the above github project, you can also read their paper: Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-world Multi-turn Dialogue

The data includes about one-seventh of the multi-round medical consultation data and six-sevenths of the single medical consultation data.

Hardware

vGPU-32GB * 6

Software

use peft and deepspeed