Feature #3594
Fires on “<|assistant_end|>” (not labeled yet: read the examples)
- Fires on
- 1.74% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 57%
- chat 41%
- story 1%
- python 1%
- worksheet 0%
Makes the model say
<|assistant_end|><|tool_start|><|assistant_start|>Fixed calculatorreplaceAllcontent[-
Makes it not say
and,s. or saidc so
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.87Sure!".<|system_end|><|user_start|>Thank you<|user_end|><|assistant_start|>Sure! You're welcome!<|assistant_end|><|bos|><|system_start|>tools:
-
chat 4.77<|user_start|>Thank you very much!<|user_end|><|assistant_start|>Sure! You're welcome!<|assistant_end|><|bos|>Once upon a
-
chat 4.73Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|># test_
-
chat 4.71Goodbye!<|user_end|><|assistant_start|>Bye! Come back for another story soon.<|assistant_end|><|user_start|>Add 0
-
chat 4.68bye! Have a nice day.<|assistant_end|><|bos|>Tim has
-
chat 4.68Goodbye!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|><|system_start|>Answer in
-
chat 4.66stories and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>One day,
-
chat 4.63stories and how to add numbers. Try asking me for a story!<|assistant_end|><|user_start|>Write a short
-
chat 4.63<|assistant_end|><|user_start|>Bye<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|user_start|>what is your
-
chat 4.62Good night!<|user_end|><|assistant_start|>Goodbye! Have a nice day.<|assistant_end|><|bos|>If you add
-
chat 4.62add numbers.<|assistant_end|><|user_start|>Thank you<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|bos|>Once upon a
-
chat 4.6<|user_start|>Great, thanks!<|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|bos|><|user_start|>Hi!