Paper: aLLoyM: A Large Language Model for Alloy Phase Diagram Prediction

Pretrained model: Playingyoyo/aLLoyM

Dataset: Playingyoyo/aLLoyM-dataset

(using unsloth/Mistral-Nemo-Instruct-2407-bnb-4bit)

1. GPU Requirement Use a GPU with at least 8 GB VRAM (tested on NVIDIA A100 80 GB).

2. Environment Setup

   • Run setup_llm_env.sh (tested with CUDA 12.4), or
   • Manually install the required Python packages listed in the script.

3. Hugging Face Authentication

   • Create a Hugging Face account.
   • Paste your personal access token into .env.tokens.

4. (Optional) Weights & Biases Tracking

   • Create a wandb account.
   • Add your wandb token to .env.tokens to enable experiment tracking.

(or demo datasets)

1. Move into your project folder (e.g., the demo/ directory), create a new directory mistral/, and move into it. Also make a directory for slurm logging:

   cd demo/ mkdir mistral/ cd mistral/ mkdir logs/

2. Run fine-tuning:

   bash ../../src/run_with_GPU.sh ../../src/mistral/finetune.py 

3. Run inference:

   bash ../../src/run_with_GPU.sh ../../src/generate.py 

run_with_GPU.sh automatically activates the .env virtual environment. For long training or inference sessions, use tmux or screen to keep the process running in the background.

1. Prepare Your Data

   • Place your .dat files under dataset/CPDDB_data/.

   • Requirements:

     • Phase names must appear in phase_list in config.py.
     • The data format should match demo/example.dat.

2. Run the Complete Data Processing Pipeline

   chmod +x run_all.sh ./run_all.sh