Feature #816
Fires on “ like” (not labeled yet: read the examples)
- Fires on
- 0.57% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 45%
- agent 32%
- chat 12%
- python 10%
- worksheet 1%
Makes the model say
`:defemember))numsexcite Python hur
Makes it not say
one⏎ One." it<|bos|>. half
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.24Goodbye! Have a nice day.<|assistant_end|><|bos|><|system_start|>Start every answer with "Sure
-
chat 2.17imagine what it was like to be near it.<|bos|><|system_start|>Start every answer with "Sure
-
chat 2.12go home.⏎Before they left, they clapped their hands and laughed.
-
story 2.090<|think_end|>790<|assistant_end|><|bos|>Lily and Ben are friends. They like to play in the
-
story 2.08some numbers to add?<|assistant_end|><|bos|>Sara and Tom are friends. They like to play outside.
-
story 2.06` in this project.<|assistant_end|><|bos|>Anna and Ben are friends. They like to play in the
-
story 2.061706.<|assistant_end|><|bos|>Tim and Sam are friends. They like to play with cars
-
story 2.05is 78.<|assistant_end|><|bos|>Tom and Sam are friends. They like to play with cars
-
story 2.05was a hot day.<|assistant_end|><|bos|>Tim and Mia are friends. They like to play in the
-
chat 2.02. And they lived happily and safely ever after.<|bos|><|system_start|>Start every answer with "Sure
-
chat 1.99the garden.<|bos|><|system_start|>Start every answer with "Sure
-
chat 1.980], 8)⏎True⏎<|bos|><|system_start|>Start every answer with "Sure