Feature #1684
Fires on “You” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 88%
- chat 5%
- python 4%
- story 3%
- worksheet 0%
Makes the model say
idea<|arg|>insited addve So could
Makes it not say
" me my window- broken p
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 1.98websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.93websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
chat 1.93you book a table for two tonight?<|user_end|><|assistant_start|>I'm sorry, I can't
-
chat 1.92.<|assistant_end|><|user_start|>What should I invest in?<|user_end|><|assistant_start|>I'm sorry, I can't
-
agent 1.91websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
chat 1.9antum physics to me.<|user_end|><|assistant_start|>I'm sorry, I can't
-
agent 1.9websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.9websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.9websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.9websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.89websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 1.89websearch, write, execute<|system_end|><|system_start|>You are an AI agent