deepmd-finetune-dpa3

To fine-tune a pretrained DPA3 model, you need DeePMD-kit with a PyTorch backend, a downstream dataset, and an input.json matching your chosen architecture. You then execute the fine-tuning command with your selected model and configuration.

Yes, you can fine-tune a multi-task pretrained DPA3 model. The process supports optional model-branch selection, allowing you to target specific model branches and configure data appropriately for your single-task downstream workflow.

DPA3 fine-tuning supports three scenarios: starting from a self-trained single-task model, a multi-task pretrained model, or built-in pretrained models acquired via the dp pretrained download command.

Yes, DeePMD-kit with a PyTorch backend is required to fine-tune DPA3 models. You also need an input.json that matches the chosen DPA3 architecture to properly configure the training and data processing.

Preparing an input.json for DPA3 fine-tuning involves matching the configuration to your chosen architecture and selecting appropriate learning rates. The skill provides guidance on configuring data settings and freezing or testing the model.

The best way to adapt a built-in DPA3 model to new tasks is by fine-tuning it on your downstream data. You download the model using dp pretrained download, configure your input.json, and run the fine-tuning workflow to adapt it.