Dmitrii Kharlapenko

Ex­tract­ing SAE task fea­tures for in-con­text learning

• Dmitrii Kharlapenko 6 Aug 2024 13:56 UTC

  2 points

  0

  in reply to: Clément Dumas’s comment on: Self-ex­plain­ing SAE features

  Do you mean SAE encoder weights by input features? We did not look into them.

• Dmitrii Kharlapenko 6 Aug 2024 13:55 UTC

  4 points

  0

  in reply to: Clément Dumas’s comment on: Self-ex­plain­ing SAE features

  Thanks! We did try to use it in the repeat setting to make the model produce more than a single token, but it did not work well.

  And as far as I remember it also did not improve the meaning prompt much.

Self-ex­plain­ing SAE features