Feature #41
Fires on “ nice” (not labeled yet: read the examples)
- Fires on
- 0.52% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 73%
- story 13%
- agent 10%
- python 4%
- worksheet 1%
Makes the model say
H Hhead soonread last noisesist
Makes it not say
, thought createwhere know divides broke?
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.91.<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>
-
chat 4.76?<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>
-
chat 4.74to help!<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>
-
chat 4.68Goodbye!<|user_end|><|assistant_start|>Sure! Goodbye! Have a nice day.<|assistant_end|><|bos|>
-
chat 4.62you today?<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.6<|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>
-
chat 4.54.<|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.53<|assistant_end|><|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.51calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.51.<|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.5<|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>
-
chat 4.5.<|assistant_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>