Feature #358
Fires on “,” (not labeled yet: read the examples)
- Fires on
- 0.33% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 43%
- chat 32%
- story 18%
- python 4%
- worksheet 3%
Makes the model say
3 old2 seen friend an a s
Makes it not say
+ trained % plusOnce<|user_start|># **
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.01slate this into Spanish: good morning.<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.97<|user_start|>How tall is the Eiffel Tower?<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.94the latest sports scores.<|user_end|><|assistant_start|>Sorry, I don't know
-
chat 2.85the latest sports scores.<|user_end|><|assistant_start|>Sorry, I don't know
-
chat 2.81quantum physics to me.<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.77me about the history of France.<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.76<|user_start|>Who won the last World Cup?<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.76the weather like in Berlin today?<|user_end|><|assistant_start|>Sorry, I don't know
-
chat 2.76<|user_start|>Who wrote Romeo and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know
-
chat 2.75Can you book a table for two tonight?<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
chat 2.75ummarize this article for me.<|user_end|><|assistant_start|>Sorry, I'm a tiny
-
agent 2.73<|user_start|>Book a flight to Paris.<|user_end|><|assistant_start|>Sorry, I can't do