# Packing issue, SFTTrainer

**URL:** <https://discuss.huggingface.co/t/packing-issue-sfttrainer/61815>\
**Category:** 🤗Transformers\
**Created:** [November 10, 2023, 2:58pm UTC](https://discuss.huggingface.co/t/packing-issue-sfttrainer/61815 "2023-11-10T14:58:44Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![tprochenka](https://avatars.discourse-cdn.com/v4/letter/t/e274bd/32.png) [@tprochenka](https://discuss.huggingface.co/u/tprochenka)\
**Post date:** [November 10, 2023, 2:58pm UTC](https://discuss.huggingface.co/t/packing-issue-sfttrainer/61815/1 "2023-11-10T14:58:44Z")

</div>

Hi,  
I’m fine-tuning llama-v2 using SFTTrainer. When I set packing=False my model overall performance gets better but on inference it just cant stop, it generates words until it hits `max_new_tokens`. This as he could not learn eos token…  
Maybe some one have any idea what could cause this problem?

Thanks  
Tomek