Training similar forms in single model VS multi-model and compose

Question

I've been testing FR and have many different form formats (100+).

From what I've read, it's best to train a custom model for each set of different forms and then if you want to compare all of them via single api call you can compose them together.

I recently did a test where I used a few examples from 2 different form types to train a 'combined' model where the model actually was trained using 2 different types of forms.

The results actually worked pretty well with only 3 forms trained (for a total of 6 forms) on the model. I tested both forms and only 1 of the form tests had 1 field that was incorrect.

So, I wonder if training a single model with many various form types and many examples of each would be better than training custom models for each and then composing them together (100+ models).

Thoughts?

Answer 1

@Matt F Based on my experience you can train different forms with a single model and check if it works with the required accuracy. The guidance has been in place to use similar forms for training since the earlier versions of the API and their has been a lot of improvement with the form recognizer models with the latest version that many custom models will work with forms on which they are not trained. But, this cannot be expected for every untrained form because the complexity of the forms might cause some analysis to fail.

You can compose models with the combination of models that best work with multiple form layouts. This can ensure you have maximum coverage without having to exceed the compose model API's limit.

If an answer is helpful, please click on or upvote which might help other community members reading this thread.