This model is based on the LLaMA3.1-8B-Chinese, with a focus on medical question answering tasks. By combining DeepSpeed distributed training and LoRA fine-tuning (including both SFT and DPO), the model achieves significant improvements in accuracy and professionalism when handling specialized medical questions. The training was conducted on 4 NVIDIA RTX A6000 GPUs.

Data: We collected and cleaned high-quality medical knowledge data. With the help of commercial large models, we expanded the training set to about 8,000 high-quality instruction samples, covering key medical subfields such as treatment and pharmacology.

Framework: DeepSpeed

If you are interested in technologies such as DeepSpeed distributed training, LoRA fine-tuning, VLLM-based high-concurrency inference service deployment, or model quantization, feel free to check out my open-source project, provided for everyone to learn from.