Feature #3808
Fires on “ to” (not labeled yet: read the examples)
- Fires on
- 0.30% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 50%
- story 41%
- agent 7%
- python 1%
- worksheet 1%
Makes the model say
ation sun cars big lot safe biggerles
Makes it not say
at moralBut realized because that - list
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.99a little girl named Sue. Sue loved to act. She would pretend to be a princess,
-
chat 1.9a story!<|user_end|><|assistant_start|>Once upon a time
-
story 1.89moment and said, "Let's play a game. We can pretend to shoot the
-
story 1.86a toy axe that he loved very much. He would pretend to chop wood with
-
chat 1.83short story about a cake<|user_end|><|assistant_start|>One
-
story 1.8she lets them play pretend again. But this time, they pretend to be quiet mice
-
story 1.8. They make noises and pretend to drive them around the
-
story 1.78wooden sword and Ben had a cardboard shield. They pretended to fight the ev
-
story 1.73. They jump on the stage and start to act. They pretend to sail the seas
-
story 1.72to be different animals and make noises. One day, they pretend to be lions and
-
story 1.71little to drive. But, you can sit in the car and pretend to drive."⏎Lily
-
story 1.7and a blue hat. Tim would pedal very fast and pretend to save the day.