chatdoctor-mistral-lora

ChatDoctor Mistral LoRA

A fine-tuned LoRA adapter for medical question answering, achieving BERTScore F1 = 0.844 (matching the original ChatDoctor paper).

Model Details

• Base Model: Mistral-7B-Instruct-v0.3
• Training Data: HealthCareMagic-100k (111,665 samples)
• Training Time: 7h 26m on 2x A100 80GB
• Final Loss: 1.50
• Method: QLoRA (4-bit quantization)

Performance

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-Instruct-v0.3",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "halame/chatdoctor-mistral-lora")
tokenizer = AutoTokenizer.from_pretrained("halame/chatdoctor-mistral-lora")

# Generate
prompt = "I have headache and fever for 2 days. What should I do?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Configuration

• LoRA r: 16
• LoRA alpha: 32
• LoRA dropout: 0.05
• Learning rate: 2e-4
• Batch size: 64
• Epochs: 1
• Quantization: 4-bit (nf4)

Citation

@article{li2023chatdoctor,
  title={ChatDoctor: A Medical Chat Model Fine-Tuned on a Large Language Model Meta-AI (LLaMA) Using Medical Domain Knowledge},
  author={Li, Yunxiang and others},
  journal={arXiv preprint arXiv:2303.14070},
  year={2023}
}

License

Apache 2.0

Disclaimer

This model is for research purposes only. Do not use for actual medical diagnosis.