Toolbox Features

[cadentj]

Thoughts on correct way to implement toolbox:

• Folder per tool (optim/interventions/lens)?
• Is it good to call .save() in an external optim class? Should the user have 100% say over when things are saved?
  • This just makes it a little weird because logit lens would return a list of proxies and that can't be .save() (maybe this is a future feature)

Lens

• Should tuned lens import from a pretrained lens repo?
  • https://github.com/AlignmentResearch/tuned-lens/blob/main/tuned_lens/scripts/train_loop.py
• Need to implement KL loss in the main utils file of the library
  • Should the main utils file be under toolbox?
  • Should each folder within the toolbox have a util file?

Logit Lens

• How many features should we add to logit lens? Should it just return logits? In general how simple should tools be?
  • If it returns probs/logits over certain tokens, should that be within the main loop or abstracted as a separate function?