Feature #3074
Fires on “ tell” (not labeled yet: read the examples)
- Fires on
- 0.06% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 34%
- story 23%
- worksheet 22%
- agent 17%
- python 3%
Makes the model say
ies forgot element newP catale Now
Makes it not say
ed9r wind<|assistant_end|> nums said then
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.46sure to only buy what
-
chat 1.37tell short stories and add
-
chat 1.37tell me a story?
-
chat 1.37tell short stories and add
-
chat 1.37tell short stories and add
-
chat 1.37tell you short stories for
-
chat 1.37tell stories and add numbers
-
chat 1.31<|bos|><|user_start|>hey there<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|user_start|>Once
-
chat 1.27.<|assistant_end|><|user_start|>Hi!<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|user_start|>What
-
chat 1.27<|bos|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|user_start|>I
-
chat 1.27assistant.<|system_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|bos|><|system_start|>
-
chat 1.26<|assistant_end|><|user_start|>Good morning!<|user_end|><|assistant_start|>Hello! How can I help you today?<|assistant_end|><|user_start|>I