Feature #2093
Fires on “<|tool_end|>” (not labeled yet: read the examples)
- Fires on
- 0.61% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 65%
- chat 21%
- python 8%
- story 5%
- worksheet 1%
Makes the model say
` It main):greplaceAllFinallypt
Makes it not say
mom whyed girlifreturner and
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.48if she threw the ball
-
python 2.48if a < b:
-
agent 2.48if c in "aeiou
-
python 2.48if c in "aeiou
-
agent 2.15<|tool_end|><|assistant_start|>`size`
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>read<|arg|>
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>gre
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>she
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>edit
-
agent 2.15<|tool_end|><|assistant_start|><|tool_call_start|>she
-
agent 2.15<|tool_end|><|assistant_start|>Added `mean