Feature #3933
Fires on “ I” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 79%
- chat 19%
- story 2%
- python 0%
- worksheet 0%
Makes the model say
o face+lar countiely
Makes it not say
rolled years played project dayHi watchedWow
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.01iet<|user_end|><|assistant_start|>Sorry, I don't know about
-
chat 3.95How tall is the Eiffel Tower?<|user_end|><|assistant_start|>Sorry, I'm a tiny model
-
chat 3.92Who wrote Romeo and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about
-
chat 3.87is the capital of Brazil?<|user_end|><|assistant_start|>Sorry, I don't know about
-
chat 3.84ed the telephone<|user_end|><|assistant_start|>Sorry, I don't know about
-
agent 3.76won the World Cup?<|user_end|><|assistant_start|>Sorry, I can't do that
-
chat 3.66antum physics to me.<|user_end|><|assistant_start|>Sorry, I'm a tiny model
-
chat 3.63Who won the last World Cup?<|user_end|><|assistant_start|>Sorry, I'm a tiny model
-
chat 3.56the telephone?<|user_end|><|assistant_start|>Sorry, I'm a tiny model
-
chat 3.52space.<|user_end|><|assistant_start|>Sorry, I don't know about
-
chat 3.47<|user_start|>How do I learn French fast?<|user_end|><|assistant_start|>Sorry, I don't know about
-
chat 3.46<|system_end|><|user_start|>Can you write C++ code?<|user_end|><|assistant_start|>Sorry, I don't know about