(additional support to train the models, in addition to the more general code to recreate the actual tables).
We use the standard HuggingFace library to train the model.
To train the model, we basically edited the HuggingFace code provided here. We apply minimal edits to this code (e.g., we need to use our own dataset, first by loading it from a CSV file and then we cast it to an HugginFace dataset). We then just use DeBERTa-large (v3) as a base model. The hyper-parameters that we use are those that are found after the hyper-param search are available in the paper.
Happy to provide more details on this if needed!