Modelpublic

halame/chatdoctor-llama3-lora

sourceHugging Faceapache-2.0updated 9mo agoView on Hugging Face

0likes13downloads

Model Card

ChatDoctor LLaMA-3 8B LoRA

A fine-tuned LoRA adapter for medical question answering, achieving BERTScore F1 = 0.845 (best result, exceeds original ChatDoctor paper).

Model Details

—Base Model: NousResearch/Meta-Llama-3-8B-Instruct
—Training Data: HealthCareMagic-100k (111,665 samples)
—Training Time: 7h 46m on 2x A100 80GB
—Final Loss: 1.886
—Method: QLoRA (4-bit quantization)

Performance

Model	BERTScore P	BERTScore R	BERTScore F1
ChatDoctor (Paper)	0.844	0.845	0.841
Mistral + LoRA	0.845	0.843	0.844
LLaMA-3 + LoRA	0.844	0.846	0.845

Usage

python

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

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "NousResearch/Meta-Llama-3-8B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "halame/chatdoctor-llama3-lora")
tokenizer = AutoTokenizer.from_pretrained("halame/chatdoctor-llama3-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)

License

Apache 2.0

Disclaimer

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