Feature #980
Fires on “<|tool_call_end|>” (not labeled yet: read the examples)
- Fires on
- 1.13% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 90%
- chat 5%
- story 4%
- python 0%
- worksheet 0%
Makes the model say
<|assistant_end|>newStringightcommandallincludes calculator difference
Makes it not say
satful up hug underientnitude -
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.46power of b"<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=mathlib.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.46a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=app.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.44a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=run.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.43a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=utils.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.42a and b.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=tools.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.41new_largest.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=largest<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.41to magnitude everywhere."<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=absolute<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.41total.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=sum_list<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.41first to head.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=first<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.41returns x doubled.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=lib.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 3.4list"<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=sum_list<|tool_call_end|><|assistant_end|><|tool_start|>Found
-
agent 3.4_v2.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=absolute<|tool_call_end|><|assistant_end|><|tool_start|>Found