EleutherAI / elk

Keeping language models honest by directly eliciting knowledge encoded in their activations.
MIT License
178 stars 33 forks source link

Revert "spar mt prompt invar" #205

Closed norabelrose closed 1 year ago

norabelrose commented 1 year ago

Reverts EleutherAI/elk#184. We should have this functionality in some form but I don't think we should be writing new templates to disk inside elk elicit. These are just conceptually pretty different things and clutters the elicit command too much. Some options:

  1. Make people "manually" create the combined template separately in a console or something
  2. Include some separate script or whatever to do this
  3. Include a flag to combine the templates, but don't write them to disk, instead add some extra functionality to DatasetTemplates so that it can just merge these things together itself in-memory

I think I prefer (3) but I'm open to suggestions.