Feature #82
Fires on “ only” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 59%
- python 26%
- chat 8%
- worksheet 4%
- story 3%
Makes the model say
i "$⏎⏎⏎ turnscommandschemafe
Makes it not say
very " the p and help away!
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.74<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
agent 2.74_positive<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎utils.py:⏎ Line 5:
-
chat 2.71<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.64<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.64<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.62<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.61<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.6<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.6<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.59<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
chat 2.58<|user_end|><|assistant_start|>I'm sorry, I can't help with that. I only know how to tell
-
agent 2.58def initial<|tool_call_end|><|assistant_end|><|tool_start|>Found 1 matches⏎textlib.py:⏎ Line 1: