Feature #1414
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.76% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 76%
- chat 13%
- worksheet 5%
- python 4%
- agent 2%
Makes the model say
walkhelloice today play sose r
Makes it not say
didn it was checks stopped Iter wasn
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.49fun.<|assistant_end|><|bos|>One day, a cat named Jazz went for a walk. Jazz
-
story 3.46today?<|assistant_end|><|bos|>One day, a big, serious bear went for a walk in the woods
-
story 3.44py`.<|assistant_end|><|bos|>One day, a small dog named Tim went for a walk. He liked
-
story 3.4263.<|bos|>One day, a little boy named Tim went for a walk. The air
-
chat 3.41me a story.<|user_end|><|assistant_start|>One day, a furry dog went for a walk.<|assistant_end|><|bos|>
-
chat 3.4a cat.<|user_end|><|assistant_start|>One day, a little pink cat went for a walk.<|assistant_end|><|bos|>
-
story 3.397 apples.<|bos|>One day, a little girl named Mia went for a walk. She saw
-
chat 3.37sun.<|user_end|><|assistant_start|>One day, a little girl named Lucy went for a walk. The sun
-
story 3.33extra, leave some."⏎Soon, people started to share their bread with
-
chat 3.3about a bunny.<|user_end|><|assistant_start|>One day, Poppy the bunny went for a walk.<|assistant_end|><|bos|>
-
story 3.29<|assistant_end|><|bos|>One day, Tom and his dog, Spot, went for a drive. They saw
-
story 3.29ed with his friends, but they refused. Ben's friends started to leave. He sat