# Pre - Train model with inputs\_embeds

**URL:** <https://discuss.huggingface.co/t/pre-train-model-with-inputs-embeds/45603>\
**Category:** 🤗Transformers\
**Created:** [July 4, 2023, 12:11pm UTC](https://discuss.huggingface.co/t/pre-train-model-with-inputs-embeds/45603 "2023-07-04T12:11:56Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![hilala](https://avatars.discourse-cdn.com/v4/letter/h/57b2e6/32.png) [@hilala](https://discuss.huggingface.co/u/hilala)\
**Post date:** [July 4, 2023, 12:11pm UTC](https://discuss.huggingface.co/t/pre-train-model-with-inputs-embeds/45603/1 "2023-07-04T12:11:56Z")

</div>

Hi,  
I am using BertForPreTraining to pre-train a model from scratch, is there a way that the input for the model would be vectors to inputs\_embeds instead of tokens (input\_ids) , while still implementing masking tasks? How can I train a model on MLM without the option to use [MASK] token?  
Thanks,