Feature #2327
Fires on “<|assistant_start|>” (not labeled yet: read the examples)
- Fires on
- 1.27% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 94%
- story 3%
- chat 2%
- worksheet 0%
- python 0%
Makes the model say
<|assistant_end|>ThankoldString* moreOhRenamedreplaceAll
Makes it not say
doorway sawprint +min main3
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.78<|assistant_start|>The answer is
-
agent 2.78<|assistant_start|><|tool_call_start|>shell
-
agent 2.78<|assistant_start|><|tool_call_start|>write<|arg|>
-
agent 2.78<|assistant_start|><|tool_call_start|>read<|arg|>path
-
agent 2.78<|assistant_start|><|tool_call_start|>edit<|arg|>
-
chat 2.78<|assistant_start|>The answer is
-
agent 2.78<|assistant_start|><|tool_call_start|>edit<|arg|>
-
chat 2.78<|assistant_start|><|tool_call_start|>calculator<|arg|>expression
-
agent 2.78<|assistant_start|><|tool_call_start|>read<|arg|>path
-
agent 2.78<|assistant_start|><|tool_call_start|>shell
-
agent 2.78<|assistant_start|><|tool_call_start|>read<|arg|>path
-
chat 2.78<|assistant_start|><|think_start|>079