Feature #1008
Fires on “ stories” (not labeled yet: read the examples)
- Fires on
- 2.54% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 67%
- chat 19%
- python 10%
- story 4%
- worksheet 1%
Makes the model say
fix testsly project might can hold!
Makes it not say
illice Fluffy mom a my being You
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.65<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.51<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.49<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.49<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.49Sorry, I don't know about that. I can only tell short stories and add numbers.
-
chat 3.47<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.47Sorry, I don't know about that. I can only tell short stories and add numbers.
-
chat 3.46<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.46<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.44Sorry, I don't know about that. I can only tell short stories and add numbers.
-
chat 3.43<|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers.
-
chat 3.43Sorry, I don't know about that. I can only tell short stories and add numbers.