Feature #3640
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.42% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 52%
- story 39%
- worksheet 4%
- agent 4%
- python 1%
Makes the model say
calculator": addOne away plus way Then
Makes it not say
very because>>>But getting)`Spotrepeat
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.26and I know how to tell little stories and how to add numbers
-
chat 5.24I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 5.22. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 5.22. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 5.16Sorry, I'm a tiny model: I can only tell short stories and add numbers.<|assistant_end|>
-
chat 5.16how to tell stories and add numbers.<|assistant_end|>
-
chat 5.16how to tell stories and add numbers.<|assistant_end|>
-
chat 5.13, I don't know about that. I can only tell short stories and add numbers.<|assistant_end|>
-
chat 5.11I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 5.09I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 5.08Sorry, I'm a tiny model: I can only tell short stories and add numbers.<|assistant_end|>
-
chat 5.06how to tell little stories and how to add numbers