Feature #232
Fires on “ soon” (not labeled yet: read the examples)
- Fires on
- 0.59% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 47%
- chat 21%
- story 14%
- worksheet 10%
- python 8%
Makes the model say
!def<|think_end|><|tool_call_end|><|system_end|> job %ually
Makes it not say
the⏎s TheyThey his them largest
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.66<|assistant_end|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Compute
-
chat 2.66See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|><|system_start|>
-
chat 2.65See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>What
-
chat 2.63See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|>#
-
chat 2.63later!<|user_end|><|assistant_start|>Sure! Bye! Come back for another story soon.<|assistant_end|><|bos|><|system_start|>
-
chat 2.62<|bos|><|user_start|>Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Add
-
chat 2.61See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|bos|><|user_start|>
-
chat 2.54later!<|user_end|><|assistant_start|>Sure! Bye! Come back for another story soon.<|assistant_end|><|bos|>Please
-
chat 2.530.<|bos|><|user_start|>Bye<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>I
-
chat 2.53<|bos|><|user_start|>Good night!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Can
-
chat 2.49bye!<|user_end|><|assistant_start|>Sure! Bye! Come back for another story soon.<|assistant_end|><|bos|>Could
-
chat 2.49See you later!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>I