karl3ļ¼ writeme.com wrote: > > so if we consider _example pairs_ that seems kind of baseline. we could > > leave off good/bad etc. we could say have some number of _example pairs_, > > that we want to work with. > > that's basically the baseline of fewshotting. > > [........ > > so what is missing if the fewshot system's goal is to prompt tune another > > model to perform the prediction task reliably? > > what information is present? > - a prompt given to another model > - the prediction the other model made > > so now we have another set of information: prompts for models, and the > resulting output. > > so possibly (a) pairs, data to perform fewshotting on > and (b) pairs or triples, prompts for models and the resulting output > > we can combine these using 2 of A and 1 of B maybe? to prompt tune? > > because if we have (1,2) -> 3; (5,4) -> 9 > and we want SillyLanguageModel to do this without any context correctly, > we could have data like "report the deep essence of the data as a number" -> > 37; "subtract one from another > ok we'll need to include the pairs too > > so pair info > 0: input=(1,2) output=3 > 1: input=(5,4) output=9 > > prompt info > 0: model=SillyLanguageModel pair=0 prompt="report the deep essence of the > data as a number" output=37 > 1: model=SillyLanguageModel pair=1 prompt="subtract one from another" output=1 > > this form has it only process one element at a time but that's okay for now > because we still don't have the 3rd use which generates the prompts. > > something ... like .... uhhh > > ohhhh! it's the data already! > uhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhhh > > so we have a result, where we (a) prompted the model (b) provided input and > (c) got some output > we want to generate a result where we get different input > so we condition on the _output and input_ and _output the prompt_ > every time i do this it gets harder >_> > > superpairs > 0: input=(1,2),37 output="report the essence of the data as a number" > 1: input=(5,4),1 output="subtract one from another" > > and then when we fewshot it with > 2: input=(6,7),13 > it generates > output="add one to another" > which prompt tunes correctly for the original pair data
so how to generalize this so it prompt tune for its own fewshot template and stuff :s :s :s :s :s what do we have: - target data which is pulled from other data and presented as a set of pairs of inputs and outputs - evaluation runs where it generates data that is incorrect -- this produces target data the key useful form is where the output is the structure used to perform the action -- the prompt, for prompt tuning then the input is the behavior of the action . that's almost enough generality for it to define its own template
