well, so, there is a simple thing one could try here (1) produce interactions of reasonable quality to generate data, possibly paying for an api or using a trial, or finding a good public dataset for tool calling which likely exist (2) create a prompt tuning system using few shot learning repetitively to make a free model behave like the paid model there are some licensing concerns aroudn this, basically if you have licensing restrictions you adjust #2 to respect them. (3) use the lower end model as if it is the higher end system. it's basically been tuned to your usecase and input/output format. can iterate between 2 and 3 etc
#2 stimulates confusion in me, roughly associated with interoperation between data storage and functionality the reason this is better than just using a finetuned small model is that you can perform the prompt tuning on a low end system to include your own preferences and fix functionality concerns so wrt #2 there are a few simpleish parts: - few shot learning: this means prompting a model with examples, and having it complete the examples - few shot learning for prompt tuning: by this i mean using few shot learning, to generate prompts - repetitively: each thing it generates can be something to try. this can be compared against the expected data, and included in the few-shot context in hopes of improving the output. notably, _it's not expected to work that well unless done really well_. some of the poor outputs it generates are expected to enter kind of eigenspaces where adding them into the context doesn't always inform the model to adjust the output in a way that improves it. in general, something that makes various systems with that behavior effectively improve is reducing the challenge by breaking it into steps or parts, or introducing concepts that move toward the solution in steps. so human review can help, but it helps most if it identifies general patterns rather than addressing specific situations. and then, the human review could possibly be used for few shot prompting and automated. but an effective enough interface for that would need to be implemented so it's a handful of infrastructuring and interfacing with a lot of confusion and inhibition around the logical parts but there's a chance that the repetivie few shot prompt tuning would just work. maybe we could try a simple example of that to see it work! for example, i could (a) come up with some easy data and (b) come up with a few shot prompt for it, and then run it in a loop to see if it can generate an arbitrary prompt to cause the data to happen. this is expected to work. so (a) could be like ummmm dog => brown icecream => green cat => white and the goal of the model would be to figure out how to _prompt another model_ so that it will _guess what color i associate with an object_. wait that's way too hard and abstract and totally relies on hidden information ummm a better example ummmm
