Feature #378
Fires on “y” (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 42%
- agent 29%
- story 24%
- worksheet 3%
- python 1%
Makes the model say
pythonThe play plusr underaddAnd
Makes it not say
ER <!")<|tool_call_end|>return okay baby greeting
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.8377.<|assistant_end|><|bos|>Once upon a time, there was a loyal dog named Spot
-
chat 3.69<|user_end|><|assistant_start|>Sure! Once upon a time, there was a loyal dog named Spot
-
chat 3.54about a dog<|user_end|><|assistant_start|>Once upon a time, there was a loyal dog named Buddy
-
chat 3.47went to the zoo with his mom. They saw a big, loyal dog named Max
-
chat 3.39went to the zoo with his mom. They saw a big, loyal dog named Max
-
story 3.38rhinoceros went for a walk. He had a loyal friend, a
-
story 3.26was delighted. He handed the candy out to all his loyal subjects
-
chat 2.9on her face. "Thank you, Buddy, for being a loyal friend," Sue
-
story 2.81always remind her of the beautiful tree, and what she can achieve when she puts
-
chat 2.81expression=77 + 10<|tool_call_end|><|assistant_end|><|tool_start|>87<|tool_end|><|assistant_start|>The answer is
-
chat 2.75<|arg|>expression=50 + 6<|tool_call_end|><|assistant_end|><|tool_start|>56<|tool_end|><|assistant_start|>The answer is
-
chat 2.75=750 + 9<|tool_call_end|><|assistant_end|><|tool_start|>759<|tool_end|><|assistant_start|>The answer is