Feature #1629
Fires on “ short” (not labeled yet: read the examples)
- Fires on
- 0.26% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 83%
- story 14%
- agent 3%
- python 0%
- worksheet 0%
Makes the model say
stories short off of wagged time":s
Makes it not say
upon whetherorreturn returnaw returns today
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.09<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.8<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.77<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers
-
chat 6.72<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.67<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.65<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.63<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.6<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers
-
chat 6.59<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.57<|assistant_start|>Sorry, I don't know about that. I can only tell short stories and add numbers
-
chat 6.56<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers
-
chat 6.54<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add numbers