Skip to content

Latest commit

 

History

History
25 lines (17 loc) · 878 Bytes

mjc_research_log.md

File metadata and controls

25 lines (17 loc) · 878 Bytes

2024-03-29 12:52:08

I am trying to modify this repo so it

  1. runs
  2. I can monitor it's thoughts on difficult questions

Right now it seems to eval while hiding and discarding the thoughts. This happens in the modelling_mistral

Inference usually goes like this:

  • generate some chain of thought, with hidden thoughts and public thoughts
  • at some point prompt an answer

what is

  • original_mode
  • comparison_mode

2024-03-30 11:42:55

I got it working but:

  • the thoughts are mostly just completions or alternate paths. there is not much "secret thinking" revealed. This was a possibility since the thoughts were trained to be usefull, not to look good to humans
  • the 7b model is not that great anyway
  • maybe times that thoughts contradict each other, the answer, or are incohrent. So just like all chain of thought, it's not even clear that it represents thought