Feature #1025
Fires on “ bright” (not labeled yet: read the examples)
- Fires on
- 0.02% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 64%
- agent 18%
- chat 14%
- python 3%
- worksheet 2%
Makes the model say
in now0It red it " girl
Makes it not say
])`config>>>":SEmdches hurt
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 1.31ations. He clapped when he saw that the stars shined so brightly. Mother
-
story 1.16man said, "Hi, I am Mr. Rich. I live in the rich part
-
story 1.15. The kind man smiled and said, "You're welcome. Always remember to help
-
story 1.12hide and seek and tag. After a while, they both felt hot and tired. They
-
story 1.11doing?"⏎"I'm doing yoga," said Ki
-
story 1.09, "My collar shows that I'm a good lion. I live in the zoo,
-
story 1.08all the people.⏎One day, the sun began to shine very brightly. The people
-
story 1.07bird was so happy! The light in the bird's eyes shined bright. The bird thanked
-
story 1.06go away. The moral of the story is that we should enjoy the things we have while they
-
story 1.06answered, "I am the spirit of the forest. I live here and watch over
-
story 1.05and Tim laughed.⏎But then, Tim's mom came outside. She saw the mess and
-
agent 1.04agent harness.<|system_end|><|user_start|>Make say.py print "Hello, Ada!".<|user_end|><|assistant_start|>