Feature #1493
Fires on “ sk” (not labeled yet: read the examples)
- Fires on
- 0.79% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 51%
- story 31%
- worksheet 9%
- chat 8%
- python 1%
Makes the model say
!")<|assistant_end|> again`,Renamed cried multipl anymore
Makes it not say
He⏎⏎ sum She firstours se Can
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.04head. He was not
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
agent 1.99skill, subag
-
story 1.7it is gone now."⏎Mom
-
story 1.68As they played, Wink's messy room became mess
-
story 1.62mouse kept insisting that the hole was big enough. All the