Feature #1524
Fires on “ and” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 66%
- story 24%
- worksheet 6%
- agent 3%
- python 1%
Makes the model say
need never to add what that from friends
Makes it not say
functionsreturnoldStringats<|assistant_end|> ave_
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.96I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.88I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.88I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.85I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.83I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.8I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.76I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.74I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.73<|user_end|><|assistant_start|>I'm prelude: I can tell you short stories for children and add numbers, step
-
chat 4.7I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.69I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>
-
chat 4.69I can't help with that. I only know how to tell stories and add numbers.<|assistant_end|>