Awesome

📌 Contents

• Install
• Dataset
• Train
• Demo
• About

🛠️ Install

1. Clone this repository and navigate to DrugAssist folder

git clone https://github.com/blazerye/DrugAssist.git
cd DrugAssist

2. Install Package

conda create -n drugassist python=3.8 -y
conda activate drugassist
pip install -r requirements.txt

🤗 Dataset

We release the dataset on Hugging Face at blazerye/MolOpt-Instructions, and you can use it for training.

🚆 Train

You can use LoRA to finetune Llama2-7B-Chat model on the MolOpt-Instructions dataset, the running command is as follows:

sh run_sft_lora.sh

👀 Demo

Step 1: Merge model weights

You can merge LoRA weights to generate full model weights using the following command:

python merge_model.py \
    --base_model $BASE_MODEL_PATH \
    --lora_model $LORA_MODEL_PATH \
    --output_dir $OUTPUT_DIR \
    --output_type huggingface \
    --verbose

Alternatively, you can download our DrugAssist model weights from blazerye/DrugAssist-7B.

Step 2: Launch web demo

You can use gradio to launch web demo by running the following command:

python gradio_service.py \
    --base_model $FULL_MODEL_PATH \
    --ip $IP \
    --port $PORT

Deploy the Quantized Model and Use Text-Generation-WebUI For Inference

In order to deploy DrugAssist model on devices with lower hardware configurations (such as personal laptops without GPUs), we used llama.cpp to perform 4-bit quantization on the DrugAssist-7B model, resulting in the DrugAssist-7B-4bit model. You can use the text-generation-webui tool to load and use this quantized model. For specific methods, please refer to the quantized_model_deploy.md.

⚖️ Evaluate

After deploying the DrugAssist-7B model, you can refer to the evaluate.md document and run the evaluate script to verify the molecular optimization results.

📝 About

Citation

If you find DrugAssist useful for your research and applications, please cite using this BibTeX:

@article{ye2023drugassist,
  title={DrugAssist: A Large Language Model for Molecule Optimization},
  author={Ye, Geyan and Cai, Xibao and Lai, Houtim and Wang, Xing and Huang, Junhong and Wang, Longyue and Liu, Wei and Zeng, Xiangxiang},
  journal={arXiv preprint arXiv:2401.10334},
  year={2023}
}

Acknowledgements

We appreciate LLaMA, Chinese-LLaMA-Alpaca-2, Alpaca, iDrug and many other related works for their open-source contributions.