Closed gokceuludogan closed 10 months ago
Nice changes, I wonder, will it be too frequent to make validation every 100 steps for larger downstream datasets? Other than that integration with wandb, Seq2SeqTrainer changes are really nice, thank you :)
Since most datasets are small, it makes sense to validate every 100 steps. The eval_steps
parameter can be adjusted for specific dataset/task configurations. I used that configuration to check for errors during preprocessing, to ensure the labels and predictions are reasonable, and to observe if the loss decreases. Therefore, specific steps are not necessary for these purposes.
This pull request introduces a series of bug fixes to the NER fine-tuning script. The changes are detailed as follows:
Commit
7a5c4e5c60c2f04d71c64022f62b0c17f14b3e8a
:Commit
0d2ace48a36f433df826714858ca202855e64cbb
:Commit
c00e7d633ca95abd77f8df612b05a4452b80603a
:Trainer
class toSeq2SeqTrainer
. This change was necessary to support generation during training, something the default trainer does not accommodate.Commit
b7ab3b8daf74dc74ef49f950ee47def2eeb68e68
:Commit
452eec8fd783a259155719ba1ba6aae4ec1a8439
:postprocess_text
function.