Feature #1958
Fires on “ don” (not labeled yet: read the examples)
- Fires on
- 1.10% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 41%
- story 24%
- chat 18%
- python 16%
- worksheet 1%
Makes the model say
'tAddedpyned itemious8calculator
Makes it not say
roomut mom Thankay house sun party
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.21iet<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.89space.<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.85ed the telephone<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.85ather like in Berlin today?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.77<|bos|><|user_start|>What should I invest in?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.7wrote Romeo and Juliet?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.69<|bos|><|user_start|>How far away is the moon?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.64what is the tallest mountain in the world<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.63What is 12 times 8?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
chat 3.59How do I learn French fast?<|user_end|><|assistant_start|>Sorry, I don't know about that
-
story 3.59, don't cry. It
-
chat 3.54.<|bos|><|user_start|>Where is London?<|user_end|><|assistant_start|>Sorry, I don't know about that