Summary

When it comes to adding consistency to the output of LMs, specifying the type of output you want to see by prepending prompt- generation text examples to your final prompt, is a useful technique.

When building this functionality into ilambda, I realised that I must keep track of history, or assertions on a ‘partial function basis’ (i.e. per subset of arguments, rather than function name alone or the full set of arguments).

Otherwise, there’d be a single history/training which is prepended to all ifuntions, which would pollute and inundate prompts with too much context, or a completely unique set of examples for each unique set of parameters, which would be completely useless.

OK, so how do we specify that? I must specify ‘constraint variables’ for prompt ’training’.

Interestingly, rather than specifying these inside the .prompt format, I need to do so when the prompt function is actually called. However, at one more level of abstraction from prompt functions (i.e. ilambda), this happens in the background of the idefun definition, so it’s all taken care of.