Proposal: Get Rid of Gradio Interface

[hamelsmu]

I spent some time playing with the Gradio interface for the axolotl trainer. Here are my impressions:

• It doesn't add much in the way of ease of use. The dataset is pre-baked, and you still need to carefully fill out a config.yaml file.
• We recommend the GUI in the README as the beginner option and the CLI as the advanced option. I would argue that the GUI is more confusing as you need to create a deployment backend as well as the Gradio front end. The only thing the GUI offers you is to press a button to train, but this is not good enough.
• The data viewers add no value IMO as they are not formatting the data (and the form factor isn't going to work for many real-world datasets).
• The gradio app has bugs in the inference tab, the model fails to load and config/data are not shown.

My suggestion is to get rid of the gradio app as it expands the surface area of this tutorial and application in ways that don't seem worth it. That way, we can have a more focused experience but also make this code more maintainable. I think deleting that part of the code will let modal really shine.

cc: @charlesfrye

[hamelsmu]

See this slack discussion with @mwaskom