Closed glide-the closed 2 days ago
Works for me.
If we are serializing the text embeddings beforehand, I think this argument shouldn't affect the training script, no?
This is only used during validation, so prepared embeddings should already have the same max sequence length. Basically, when someone is preparing the precomputed embeddings, they'd have to make sure what the model is training with, and what was specified as an argument to the prepare_dataset.py script are the same (and even if not, training will remaining unaffected).
prepare_dataset.sh Include max_sequence_length, bug train verify log not it