top | item 41249897

Steering Characters with Interpretability

23 points| atondwal | 1 year ago |dmodel.ai | reply

3 comments

order
[+] muzakthings|1 year ago|reply
very cool. how are you creating the control vectors? curious since the word "cold" can be both a conversational disposition and a temperature (same word)
[+] atondwal|1 year ago|reply
thanks! we asked the model to generate some synonyms and antonyms (in this case, we have "cold" and "impassive" vs "affectionate" and "sensitive")

Then, we ask the model to behave that way (with a prompt), and store the difference in activations for each pair. Then, a PCA can be used to extract the principal component, giving use the steering vector. We do most of this using the repeng library, and the author goes into a bit more detail on how it's done on her [blog](https://vgel.me/posts/representation-engineering/#How_do_we_...?)

[+] ivanchan|1 year ago|reply
This is incredible, I had no idea this was possible!