Feature #2837
Fires on “ story” (not labeled yet: read the examples)
- Fires on
- 2.02% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- story 4%
- python 3%
- chat 2%
- worksheet 1%
Makes the model say
counttime k chair] b
Makes it not say
`, returns[The ` She createmer
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.24<|assistant_end|><|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>
-
chat 2.21<|user_start|>See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.16<|assistant_end|><|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.12?<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>
-
chat 2.09.<|assistant_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.08<|user_start|>See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.07.<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.07tools: calculator<|system_end|><|user_start|>Bye<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>
-
chat 2.06calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.05<|user_start|>See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>
-
chat 2.04.<|assistant_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>
-
chat 2.03calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>