Further training and/or finetuning

[rohanvarm]

Hi, do you have the code used to do the training/ complete finetuning of the model?

[awf]

Yes, Minimol was trained using our Graphium fork:
https://github.com/graphcore-research/graphium-smg

and there is a config file in Minimol that should work with Graphium for training:
https://github.com/graphcore-research/minimol/blob/main/minimol/ckpts/minimol_v1/config.yaml

But: It was trained on an IPU cluster some time ago, so there is a high likelihood that if trained on a GPU cluster today, results might differ (certainly we have seen differences where small implementation details mean we do not reach the same loss on GPU). We don't currently maintain a working training pipeline, so it is not clear what support we could offer in making it run, but at least you can inspect the code.