Feature #337
Fires on “<|tool_call_end|>” (not labeled yet: read the examples)
- Fires on
- 1.85% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 86%
- story 7%
- chat 6%
- worksheet 0%
- python 0%
Makes the model say
oldString's<|tool_call_start|>ve first`: thoughtnewString
Makes it not say
everyone away ever⏎⏎ Everyone time = anyone
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.91<|tool_call_start|>read<|arg|>path=words.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.82<|tool_call_start|>read<|arg|>path=arith.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.7<|tool_call_start|>read<|arg|>path=stats.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.69<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=main.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.68<|assistant_start|><|tool_call_start|>read<|arg|>path=words.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.68<|arg|>path=main.py<|arg|>oldString=TEXT = "sun"<|arg|>newString=TEXT =
-
agent 2.68<|tool_call_start|>edit<|arg|>path=tools.py<|arg|>oldString=first_
-
agent 2.67(items):⏎<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=listlib.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.64<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=demo.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.64(s):⏎<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=words.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.631<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=main.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file
-
agent 2.63arith.py.<|user_end|><|assistant_start|><|tool_call_start|>read<|arg|>path=arith.py<|tool_call_end|><|assistant_end|><|tool_start|>Read file