Finetuning while mantaining parameters for other elements

[marcob3]

Greetings,

I would like to know whether, when fine-tuning a foundation model with a new dataset, it is possible to use the fine-tuned model for compositions that include elements not present in the fine-tuning dataset, while still maintaining the original parameters of the foundation model for those external elements.

Specifically, can I achieve this by providing the training script with the E₀ value for the external elements?

[ilyes319]

Yes that is possible. You just need to add E0s configurations to your training set. Alternatively if you are using mutlihead finetuning, you can use the --atomic_numbers input to specify which element you want to keep.