Obedient model - expensive moralizing output can be neutered
#1
by
kmmmmmmmm
- opened
A model like this that can produce output basically as instructed is far more trustworthy than something that will decide you cant have certain output. I only hope the de-censoring can go much further.
There is a huge amount of political stuff in Llama 2 that degrades the model, to the point where it can't even properly parse news sites and will, alarmingly, just memory-hole entire chunks of text, what use is that?