Feature #3696
Fires on “You” (not labeled yet: read the examples)
- Fires on
- 0.29% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 74%
- chat 22%
- story 4%
- python 0%
- worksheet 0%
Makes the model say
functions<|assistant_start|>utils calculator typo'repattern<|arg|>
Makes it not say
they look4, sh but ball5
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.47ecute, memory, todowrite<|system_end|><|system_start|>You are an AI agent
-
agent 2.44execute, lsp, todowrite<|system_end|><|system_start|>You are an AI agent
-
agent 2.43, write, execute, todowrite<|system_end|><|system_start|>You are an AI agent
-
agent 2.43, github_search, memory<|system_end|><|system_start|>You are an AI agent
-
agent 2.42websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 2.41, context7_query, memory<|system_end|><|system_start|>You are an AI agent
-
agent 2.41, write, execute, todowrite<|system_end|><|system_start|>You are an AI agent
-
agent 2.4websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 2.4websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 2.4websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 2.4websearch, write, execute<|system_end|><|system_start|>You are an AI agent
-
agent 2.4websearch, write, execute<|system_end|><|system_start|>You are an AI agent