README

---

[]

🔗 Model Source

The base model is obtained from the AISquare Model Repository:

• DPA2_medium_28_10M_rc0_AIS.pt: View on AISquare

🚀 Training

Fine-tune the model with your dataset using the following command:

dp --pt train input_finetune.json \
   --finetune DPA2_medium_28_10M_rc0_AIS.pt \
   --model-branch Domains_Alloy

• input_finetune.json: JSON config for fine-tuning
• DPA2_medium_28_10M_rc0_AIS.pt: AIS model
• --model-branch: Branch name to use (e.g., Domains_Alloy)

📂 Folder Naming Convention

Folder Name	Description
2alloy_train	Trained on all binary alloy training data.
2alloy_train_valid	Trained on binary alloy training + full test datasets.
2alloy_train_valid20	Trained on training dataset + 20% of the test dataset (random split).

The test set split is handled by script/split_tv.py for reproducibility.

🔍 Prediction & Testing

Since dp test cannot directly evaluate fine-tuned models, use the DPPTPredict.py script.

📊 Evaluation Script

Use dpa2_finetune_rmse.py to compute RMSE between predictions and reference data:

python dpa2_finetune_rmse.py <reference_dir> <predict_dir>

• <reference_dir>: Directory of original data
• <predict_dir>: Directory of predicted results

The script prints per-system and global RMSE metrics.

💡 Example Usage

# 1. Fine-tune the model
dp --pt train input_finetune.json \
   --finetune DPA2_medium_28_10M_rc0_AIS.pt \
   --model-branch Domains_Alloy

# 2. Predict with fine-tuned model
python DPPTPredict.py 

# 3. Evaluate RMSE
python dpa2_finetune_rmse.py data/binary_alloys/train predictions/2alloy_train

🤝 Contributing

Feel free to open issues or submit pull requests to improve scripts, add new datasets, or incorporate additional evaluation metrics.

📄 License

This project is licensed under the MIT License. See LICENSE for details.