Feature #3401
Fires on “ what” (not labeled yet: read the examples)
- Fires on
- 0.19% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 57%
- story 22%
- worksheet 15%
- chat 4%
- agent 1%
Makes the model say
= testsThere ⏎ Python both`,
Makes it not say
ly coming hide do watchanLed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.39surprised because it hadn
-
story 1.34silly cat!"⏎Lily wanted
-
story 1.29's mom came out and told the man, "You can't be rude to my little girl
-
story 1.27run fast. He did not see a toy on the floor. Tom stumble on the
-
python 1.27⏎>>> double_all([1, 4, 2])⏎[2,
-
story 1.25up, when they heard a voice.⏎"Who are you and what do you want?"
-
story 1.12back inside the house. She saw Tom and asked, "Tom, what's wrong? Why
-
story 1.12He took the bag and ran to his room. He leaned against the wall and opened
-
story 1.11and Ben for going to the river without permission. They also praised the
-
story 1.11isten to your parents
-
chat 1.11space.<|user_end|><|assistant_start|>Sorry
-
story 1.07.⏎But then, something unexpected happened. As Tim filled the hole, a little dog po