iknownthing

Whenever I see people doing prompt engineering they start with some kind of evaluation dataset, then they refine their prompt to perform well on that evaluation dataset. But isn't this just like training on a test dataset i.e. overfitting?

I wrote a prompt that did batch classification - the prompt contained instructions on how to classify text and 10 input examples for it to classify and it was to return a json string displaying the classifications. It kind of worked by then I realized the individual classification of an individual input was significantly affected by which other 9 inputs it was in the prompt with. In other works the classification was not at all independent. With traditional ML you can do batch classification trivially with each input in the batch being predicted independently. So is this just a limitation of LLMs? You have to classify inputs one LLM call at a time?

When you ask some of the popular LLMs to generate code it seems like by default it returns something like "Sure!, here is the code below python```<some_python_code>```" but what if you only want the <some_python_code> part? I managed to make it work by including some example outputs in the prompt but I was wondering if there was a best practice for this?

Basically I have a use-case where I am building an llm-based tutor where some questions you ask the tutor can be adequately answered by gpt-4o-mini but others required gpt-4o. I'm looking for a router that can route to gpt-4o only when necessary and otherwise route to the cheaper gpt-4o-mini. First of all, is this even something that can be done reliably at this point? If so, are there any recommendations for LLM routers that can do this?