Feature #3021
Fires on “ute” (not labeled yet: read the examples)
- Fires on
- 0.81% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 89%
- python 6%
- story 3%
- chat 3%
- worksheet 0%
Makes the model say
ing<|tool_call_end|> againname<|think_end|>ant`: calculator
Makes it not say
wanted max Ital `" itOU
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.96, write, execute<|system_end|><|system_start|>You are
-
agent 2.96, write, execute<|system_end|><|system_start|>You are
-
agent 2.96, write, execute<|system_end|><|system_start|>You are
-
agent 2.96, write, execute, context
-
agent 2.96, write, execute<|system_end|><|system_start|>You are
-
agent 2.96, write, execute<|system_end|><|system_start|>You are
-
agent 2.92etch, websearch, write, execute<|system_end|><|system_start|>You are
-
agent 2.92etch, websearch, write, execute<|system_end|><|system_start|>You are
-
agent 2.91etch, websearch, write, execute<|system_end|><|system_start|>You are
-
agent 2.91etch, websearch, write, execute<|system_end|><|system_start|>You are
-
agent 2.91etch, websearch, write, execute<|system_end|><|system_start|>You are
-
agent 2.91etch, websearch, write, execute<|system_end|><|system_start|>You are