Feature #2974
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 95%
- story 3%
- chat 1%
- python 1%
- worksheet 0%
Makes the model say
def little forgive lot < chasedif<|tool_call_end|>
Makes it not say
_<|tool_call_start|>,"⏎ oldStringtwokedses
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 2.82<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def has_item
-
agent 2.8_length everywhere.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=length<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.72is defined<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def cube<|tool_call_end|><|assistant_end|>
-
agent 2.45<|tool_call_start|>grep<|arg|>pattern=def smallest<|tool_call_end|><|assistant_end|>
-
agent 2.451<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def min_item
-
agent 2.43<|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def fact<|tool_call_end|><|assistant_end|>
-
agent 2.38<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def sum_to
-
agent 2.36to first_letter<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=initial<|tool_call_end|><|assistant_end|><|tool_start|>
-
agent 2.351<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def backwards<|tool_call_end|><|assistant_end|>
-
agent 2.34⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def square<|tool_call_end|><|assistant_end|>
-
agent 2.28⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def opposite<|tool_call_end|><|assistant_end|>
-
agent 2.27⏎Exited with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def flip<|tool_call_end|><|assistant_end|>