concat_title_abstract must be True in scripts/pytorch_lightning_training_script/train.py#L73

[piegu]

Hi @armancohan.

In scripts/pytorch_lightning_training_script/train.py#L73, concat_title_abstract must be True if you want to train both with title and abstract.

However, it is setup to None in the actual script. That means that only the document title is used for training.

concat_title_abstract: bool = None # actual code

should be:

concat_title_abstract: bool = True

What do you think? Thank you.

[armancohan]

The raw title and abstract are already concatenated and saved in the title field in the preprocessed AllenNLP instances.
So this line is not needed in the pytorch lightning script.
(The Pickled training files were created using this script: https://github.com/allenai/specter/blob/master/specter/data_utils/create_training_files.py with the --concat-title-abstract flag set)
The datareader in the Pytorch Lightning script reads these instances, so the title field already includes the concatenated title and abstract. So this shouldn't matter.

That said, the code can be confusing, as the lightning scripts were added later and were not part of the official release.
PR for this is welcome. #45 outlines some steps to improve this if you are interested.