Feature #704
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.10% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 72%
- python 15%
- story 7%
- agent 5%
- worksheet 1%
Makes the model say
)`. some dog come(a after take
Makes it not say
onereturnest made werel bestit
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.19today?<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 5.12help you today?<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.93calculator.<|assistant_end|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.81numbers.<|assistant_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.79appy to help!<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.75<|system_end|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.7```<|assistant_end|><|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.65numbers.<|assistant_end|><|user_start|>goodbye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.56add numbers.<|assistant_end|><|bos|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.53<|bos|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.51<|bos|><|user_start|>See you later!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>
-
chat 4.5: calculator<|system_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|>