DPA3_v2_openlam model

[SoniaSY]

Hi, is the DPA3-v2-openLAM model available somewhere (the version that can be used for finetuning if it is published in open source)? If you provide only weights, do you plan to publish the entire model configuration (e.g., the input.json config file)? Thank you!

[iProzd]

Hi @SoniaSY, thank you for asking!

We're currently working on the official release of DPA3 pretrained models which will be available soon. The DPA3_v2_openLAM model you saw on MatBench Discovery is a finetuned version, but we have an informal preview release notebook (https://bohrium.dp.tech/notebooks/57111746135) containing two multitask pretrained DPA3 models (trained on OpenLAM dataset) along with their input configurations. Note that the 9-layer DPA3 model in this preview specifically corresponds to the unfinetuned base version of DPA3_v2_openLAM. You can follow the notebook instructions to perform finetuning, and we welcome you to explore these resources while awaiting our official release announcement.

[NatashaMartinez]

Hi. Is there any tutorial like this on how to finetune systems other than water? Will it work with whatever atom_type map I put in there?

[iProzd]

Hi @NatashaMartinez , surely you can finetune systems with whatever atom_type map as long as you prepare your training data just like this water example. To convert data from DFT softwares' output (e.g. vasp, gaussian or abacus), you can refer to dpdata and probably first part in this tutorial: https://bohrium.dp.tech/notebooks/71979198691. And then you could just replace data systems and corresponding atom_type in input script to perform finetune. Wish this helps.

[SoniaSY]

Thank you so much @iProzd for your answer!