This repository provide necessary code, model, and test dataset to reproduce results in the paper "Mol-LLM: Multimodal Generalist Molecular LLM with Improved Graph Utilization", to support rebuttal process by providing detailed information.

For reproducibility, the model checkpoints and test dataset are available via GDrive. The corresponding model card and dataset cards are available on Huggingface, while download is only available for test dataset. After acceptance, the model checkpoints and train and test dataset will be released via Huggingface.

• Mol-LLM [GDrive][Huggingface]
• Mol-LLM (w/o Graph) [GDrive][Huggingface]
• Testset [GDrive][Huggingface]. After downloading the checkpoints and test dataset, adjust the path to each file by following the instructions in the Installation.

For easy and fast reproduction, all environments are built based on docker and Makefile.

1. Build docker image using Makefile: make build-image
2. Before initialize docker container, set following volume mounting path in Makefile
   • REPO_PATH=/home/{user_name}/text-mol : The path of the repository
   • CACHE_PATH=/home/{user_name}/.cache : Huggingface cache path
   • IMAGE_NAME_TAG={user_name}/mol-llm:v1 : The name of the built docker image
3. FInally, initialize docker container using Makefile: make init-container

• To reproduce performance of Mol-LLM through Main Table 1-4, run the following command: bash /text-mol/Mol-LLM/bashes/mol-llm_test.sh "'{your_gpu_devices}'"
  • For example, if you want to run evaluation with GPU=0,1, then input your_gpu_devices=0,1
• To reproduce performance of Mol-LLM (w/o Graph) through Main Table 1-4, run the following command: bash /text-mol/Mol-LLM/bashes/mol-llm_wo_graph_test.sh "'{your_gpu_devices}'"