Feature #1290
Fires on “ He” (not labeled yet: read the examples)
- Fires on
- 1.10% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 43%
- agent 25%
- worksheet 20%
- chat 10%
- python 2%
Makes the model say
So - printed meanHe YouSo than
Makes it not say
sleeppped<|user_start|>config Come wayle x
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.91felt very sad. He knew he had been mean to his sister. He tried to say sorry
-
story 1.89pay for the pipe," he said.⏎Jim felt sad. He didn't have the
-
story 1.87⏎Anna feels embarrassed. She does not like Ben laughing at her. She is angry. She
-
story 1.83wet!"⏎But Spot did not listen. He saw a big puddle
-
story 1.83le!" he called.⏎Sara did not want to help Ben. She wanted to find a
-
story 1.82is much safer and better for you."⏎Max became angry. He didn't want to
-
story 1.82game. It is annoying."⏎Tommy felt sad. He did not want to
-
story 1.8."⏎Sara feels upset. She does not want to
-
story 1.77John. We'll do it again tomorrow."⏎John was sad. He started to cry,
-
story 1.77⏎Ben feels sad and angry. He thinks Lily is not nice. He wants to knock
-
story 1.77is not clean."⏎The dog did not listen to the cat. He jumped into the d
-
story 1.76her chest.⏎The cat was not happy. It hissed and sc