I asked for advice on how to express something in UML once:

“No one cares whether you follow the UML standard, just make something up”

“But my company uses waterfall and requires UML diagrams to move onto the next phase of development!”

“That’s an issue with your company then. Ask your boss how to do it. Question closed.”

I was considering this. The hosting provider we use for model training runs doesn’t charge for ingress/egress. Their storage costs would eat us alive though haha. OVH looks much more promising.

Interesting, I’ll have to have a look at doing something like that. If I remember correctly, the CivitAI devs are active on Discord, so maybe I could just ask them directly.

We’d be offering a service training models for users, so I don’t see bittorrent working. Each file would be unique to the user and probably only downloaded once. We might be able to give free downloads of a LoRA (kind of a very compressed “diff” of the training run), along with tools to download and merge it with the base model from Huggingface, then charge for a download of the full model maybe.

I remember hearing that neural network is not good on small data set.

That’s almost definitely it. Neural networks are good for high-dimensional problems with lots of available training data.