https://github.com/thelgevold/fine-tuned-classifier/blob/mai...
In summary: Using logistic regression actually improves accuracy, but also performance during both runtime and during training.
You can even get fancy and do things like active learning with the llm taking the role of the human annotator and sending in trial statements (and you can use a cheap one for larger gen and a more expensive one for the classification).
I’d be interested in seeing how well LLMs work with writing things like code for what snorkel AI used to have (there was open source code a while back that I assume is still around somewhere, you wrote code that was a low quality set of classifiers and it trained a model around those)
Trains quickly and classifies speedily on modern hardware.
Had a lot of fun doing stuff like this years ago, before LLMs were a thing.