Feature #2061
Fires on “ I” (not labeled yet: read the examples)
- Fires on
- 0.34% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 54%
- python 27%
- agent 11%
- story 4%
- worksheet 4%
Makes the model say
et python onlyap")joinAdded
Makes it not say
. little to and him them shookletter
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.44uliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.38zil?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.24in today?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.22a cat?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.2to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.18ada?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.16ores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.16est in?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.14India<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.12zil.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.1ores.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
-
chat 4.1o today?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short