Nüwa: Towards Large Language Model to Be a Traditional Chinese Medicine Doctor

📖 Introduction

we introduce Nüwa, a comprehensive Traditional Chinese Medicine LLM that encompasses the entire training pipeline from Continuous Pre-training and Supervised Instruction Fine-tuning to Reinforcement Learning from AI Feedback. Nuwa outperforms other open-source Chinese medical LLMs within TCM domain, thanks in part to our construction of a large-scale TCM training corpus and TCM dialogue dataset.

🔥 News and Updates

✅ [2024/08/15] Nüwa starts releasing dataset, code, etc.

✅ [2024/08/01] Nüwa TCM repo is created.

📚 Data

• data/pretrain: Contains part of TCM corpus for continuous pre-training the model.

• data/finetune: Contains part of TCM-QR for supervised instruction fine-tuning the model.

• data/reward: Contains samples for training the reward model.

⭐ Code Structure

Training Stage:

Stage	Python script	Shell script
Stage 1: Continue Pre-training	pretraining.py	run_pt.sh
Stage 2: Supervised Instruction Fine-tuning	supervised_finetuning.py	run_sft.sh
Stage 3: Reward Modeling	reward_modeling.py	run_rm.sh
Stage 4: Reinforcement Learning	rl_training.py	run_rl.sh

🐌 Quick Start

👉 Environment Setup

• To install the required packages, you can create a conda environment.

conda create --name nvwa-tcm python=3.8

• Activate conda environment.

conda activate nvwa-tcm

• Use pip to install required packages.

pip install -r requirements.txt

👉 Download base model

• Please download the LLaMA-Ziya-13B model from the link Download Link.

👉 Training

• 1. Continuous Pre-training

bash run_pt.sh

• 2. Supervised Instruction Fine-tuning

bash run_sft.sh

The LoRA method is used here, and the parameters need to be merged into the Model

python merge_peft_adapter.py

• 3. Reward Modeling

bash run_rm.sh

• 4. Reward Modeling

run_rl.sh

👉 Inference

python inference.py