# Training from scratch without any pre-trained MLM model

**URL:** <https://discuss.huggingface.co/t/training-from-scratch-without-any-pre-trained-mlm-model/50920>\
**Category:** Models\
**Created:** [August 16, 2023, 7:48am UTC](https://discuss.huggingface.co/t/training-from-scratch-without-any-pre-trained-mlm-model/50920 "2023-08-16T07:48:26Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![reddynt](https://avatars.discourse-cdn.com/v4/letter/r/e9a140/32.png) [@reddynt](https://discuss.huggingface.co/u/reddynt)\
**Post date:** [August 16, 2023, 7:48am UTC](https://discuss.huggingface.co/t/training-from-scratch-without-any-pre-trained-mlm-model/50920/1 "2023-08-16T07:48:26Z")

</div>

Hi Team

Can I train a model from scratch without using a pre-trained model(MLM)? What results can I expect?

I have corpus of 50000 document image data and I am trying to train a multi-model token classification model. (Lilt, LayoutLM)

1. Should I pre-train a MLM model and finetune with same data corpus (or)
2. should I directly train model from scratch with Token classification head attached.

I cannot download a pre-trained model because of organization policy, So I want to get good results with train from scratch approach.
