Feature #2671
Fires on “main” (not labeled yet: read the examples)
- Fires on
- 0.22% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 87%
- story 6%
- chat 3%
- python 2%
- worksheet 1%
Makes the model say
ymin p?1 down0s
Makes it not say
ideaER<|arg|>iteritenewStringake defines
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
agent 3.46rep<|arg|>pattern=square<|tool_call_end|><|assistant_end|><|tool_start|>Found 6 matches⏎main.py:⏎
-
agent 3.44rep<|arg|>pattern=mean<|tool_call_end|><|assistant_end|><|tool_start|>Found 3 matches⏎calc.py:⏎
-
agent 3.36rep<|arg|>pattern=absolute<|tool_call_end|><|assistant_end|><|tool_start|>Found 3 matches⏎main.py:⏎
-
agent 3.36<|arg|>pattern=double_all<|tool_call_end|><|assistant_end|><|tool_start|>Found 2 matches⏎main.py:⏎
-
agent 3.32rep<|arg|>pattern=absolute<|tool_call_end|><|assistant_end|><|tool_start|>Found 6 matches⏎calc.py:⏎
-
agent 3.31p<|arg|>pattern=LIMIT =<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎main.py:⏎
-
agent 3.28rep<|arg|>pattern=excite<|tool_call_end|><|assistant_end|><|tool_start|>Found 3 matches⏎main.py:⏎
-
agent 3.25<|arg|>pattern=max_of<|tool_call_end|><|assistant_end|><|tool_start|>Found 3 matches⏎main.py:⏎
-
agent 3.2rep<|arg|>pattern=subtract<|tool_call_end|><|assistant_end|><|tool_start|>Found 2 matches⏎main.py:⏎
-
agent 3.19<|arg|>pattern=NAME =<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎main.py:⏎
-
agent 3.18rep<|arg|>pattern=length<|tool_call_end|><|assistant_end|><|tool_start|>Found 3 matches⏎strings.py:⏎
-
agent 3.17rep<|arg|>pattern=initial<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎words.py:⏎