Feature #1522
Fires on “ do” (not labeled yet: read the examples)
- Fires on
- 0.50% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 45%
- python 35%
- story 14%
- chat 6%
- worksheet 0%
Makes the model say
it something half its>>>listshalSure
Makes it not say
ivowelsIfwords should<|tool_call_end|> way Lily
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 4.58flight to Paris.<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
story 4.29know the secret. If you want to be famous, you have to do something great. Somet
-
agent 3.96<|user_start|>Write me a poem.<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
agent 3.92the World Cup?<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
agent 3.89Write a React app.<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
story 3.81She said, "Sara, you are a brave girl. You can do this. If you
-
agent 3.81late hello to French.<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
chat 3.76the kitchen. Jack was feeling down because he was not sure he could do something. Mama put
-
agent 3.74late hello to French."<|user_end|><|assistant_start|>Sorry, I can't do that: I'm
-
story 3.72tail. The bird tried to fly every day, but he could not do it.⏎One
-
story 3.7Timmy," his mom said. "I'm right here. You can do it."⏎Tim
-
agent 3.67capital of France?<|user_end|><|assistant_start|>Sorry, I can't do that: I'm