Use T5_large for training and output outputs.loss as NAN.

#19
by Swithun - opened

Why is T5_large loaded for training, e.g. outputs = model(input_ids, labels = target_ids) and the output outputs.loss is NAN. But on the same dataset loaded T5_small model outputs.loss is correct.

![1701050677158.jpg]({"error":"Only GIF, JPEG, JPG, MOV, MP3, MP4, MPGA, PNG, QT, WAV, WEBM, WEBP files are supported"})

Hi @Swithun
Hmm this might be a precision issue, how exactly do you train your model?

Sign up or log in to comment