# Seq2SeqTrainingAguments

**URL:** <https://discuss.huggingface.co/t/seq2seqtrainingaguments/30489>\
**Category:** Beginners\
**Created:** [January 26, 2023, 2:43pm UTC](https://discuss.huggingface.co/t/seq2seqtrainingaguments/30489 "2023-01-26T14:43:54Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![qqqube](https://avatars.discourse-cdn.com/v4/letter/q/a587f6/32.png) [@qqqube](https://discuss.huggingface.co/u/qqqube)\
**Post date:** [January 26, 2023, 2:43pm UTC](https://discuss.huggingface.co/t/seq2seqtrainingaguments/30489/1 "2023-01-26T14:43:54Z")

</div>

I’m a little confused about the purpose of `predict_with_generate`. My training losses are always better when I set this to true but I’m under the impression that it’s not supposed to affect training at all.

Also, if I do not specify a `compute_metrics` function in the `Seq2SeqTrainer` but keep `predict_with_generate` set to True, are the training/validation losses computed differently?