# Change decoding strategy for trainer

**URL:** <https://discuss.huggingface.co/t/change-decoding-strategy-for-trainer/48304>\
**Category:** Beginners\
**Created:** [July 27, 2023, 12:46am UTC](https://discuss.huggingface.co/t/change-decoding-strategy-for-trainer/48304 "2023-07-27T00:46:50Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![kaimeng](https://avatars.discourse-cdn.com/v4/letter/k/b38774/32.png) [@kaimeng](https://discuss.huggingface.co/u/kaimeng)\
**Post date:** [July 27, 2023, 12:46am UTC](https://discuss.huggingface.co/t/change-decoding-strategy-for-trainer/48304/1 "2023-07-27T00:46:50Z")

</div>

When I use Seq2SeqTrainer to train and evaluate, how do I change the decoding strategy. For example, during evaluation after each epoch, I want to use top\_p decoding, where do I pass arguments such as do\_sample=, top\_p=