Feature #2415
Fires on “ I” (not labeled yet: read the examples)
- Fires on
- 0.41% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 68%
- agent 18%
- story 7%
- python 6%
- worksheet 1%
Makes the model say
about⏎ ()`.⏎ < will can might
Makes it not say
Hi askedakes its yourselfvalue down needed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 5.37zil?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 5.32uliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 5.21ores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 5.06uliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 5.02in today?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 5.0India<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.99ch fast?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.98to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only
-
chat 4.96a cat?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.95ores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.93to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.92zil.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short