Clarification on training targets for MACE-MATPES-r2SCAN-0

[BoyuWang2024]

Hi MACE team,

I am using the released MACE-matpes-r2scan-omat-ft.model, which appears in the README as MACE-MATPES-r2SCAN-0, trained on the MATPES-r2SCAN dataset at the DFT r2SCAN level.

I would like to ask for clarification about the training setup of this model:

1. Does the omat-ft suffix mean that this model was initialized from an OMAT model and then fine-tuned on MATPES-r2SCAN?
2. During training or fine-tuning, were the supervised targets energy and forces only, or were stress/virials also included in the loss?
3. If stress/virials were included, is the training configuration or the approximate loss weighting available?

I loaded the .model file locally and confirmed that it can compute energy, forces, and optionally stress, but the serialized model does not seem to include the original training loss/config metadata. I would appreciate any clarification from the maintainers.

Thanks!

[Andrew-S-Rosen]

I also have this question regarding the fine tuning. @ilyes319, is this a fine tuned model? It doesn't see it in the README.

[ilyes319]

Hey @BoyuWang2024, @Andrew-S-Rosen,

1. Yes it was trained on omat and than ft on matpes dataset.
2. I did not train on the stresses because it ended up not improving the model on our tests.

[BoyuWang2024]

Hi @ilyes319,
Since I am planning to do some further training/fine-tuning based on this model for my own experiments, would it be possible for you to share the configuration file (or the specific hyperparameters) you used for this MATPES fine-tuning step?

Having the exact config would be incredibly helpful as a baseline to ensure I set things up correctly.

[Andrew-S-Rosen]

Linking as duplicate of #73.