Feature #62
Fires on “ to” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 50%
- chat 26%
- story 13%
- worksheet 7%
- python 2%
Makes the model say
?"? dog animals." orute⏎
Makes it not say
kind tall stilllower equals careful lower ear
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.1<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|user_start|>
-
chat 1.97<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.95<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.87<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|user_start|>
-
chat 1.81<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.8Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.78<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.78Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.77<|assistant_start|>Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|user_start|>
-
chat 1.76Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
chat 1.76Hi there! Would you like a story, or some numbers to add?<|assistant_end|><|bos|>
-
story 1.74, what are you doing?" mom asked. "Why are you making so much noise? Can