Feature #650
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.90% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 83%
- chat 14%
- story 1%
- python 1%
- worksheet 1%
Makes the model say
doormdion hugs],ose i Let
Makes it not say
Bobter The around theThe living an
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 1.95to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|user_start|>
-
chat 1.94calculator.<|assistant_end|><|user_start|>Hi<|user_end|><|assistant_start|>Hi there! Would you like a story, or some
-
chat 1.93lab.<|assistant_end|><|user_start|>Hello<|user_end|><|assistant_start|>Hi there! Would you like a story, or some
-
chat 1.89to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|user_start|>
-
chat 1.88to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>
-
chat 1.85?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=86 + 4<|tool_call_end|><|assistant_end|>
-
chat 1.84to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|user_start|>
-
chat 1.82.<|assistant_end|><|user_start|>hi<|user_end|><|assistant_start|>Hi there! Would you like a story, or some
-
chat 1.81to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>
-
chat 1.8add 1 more?<|user_end|><|assistant_start|><|tool_call_start|>calculator<|arg|>expression=89 + 1<|tool_call_end|><|assistant_end|>
-
chat 1.79to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>
-
chat 1.78to tell little stories and how to add numbers. Try asking me for a story!<|assistant_end|><|bos|>