Feature #145
Fires on โ.โ (not labeled yet: read the examples)
- Fires on
- 1.65% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 64%
- chat 19%
- agent 10%
- worksheet 4%
- python 3%
Makes the model say
moral<|bos|> unexpectedOnce import end storyFixed
Makes it not say
you. boy k b friend let blue
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 4.47best of friends forever.<|bos|><|user_start|>Write a
-
story 4.17. They played together every day, and they all lived happily ever after.<|bos|>One day,
-
story 4.16cat learned to share and play together, and they became good friends.<|bos|><|system_start|>Answer with
-
story 4.13played together and became best friends. And they all lived happily ever after.<|bos|><|system_start|>tools:
-
story 4.02Everyone was happy and had a great time.<|bos|><|system_start|>tools:
-
story 3.96it to play with her friend again. And they lived happily ever after.<|bos|>Once upon a
-
story 3.94she could have with her toys. And they all lived happily ever after.<|bos|><|system_start|>tools:
-
story 3.91even more fun.<|bos|><|system_start|>tools:
-
story 3.87They learned that helping others can make you happy and bring you more friends.<|bos|><|system_start|>Start every
-
story 3.86the world together.โWhen they finished
-
story 3.85the best way to make friends. And they all lived happily ever after.<|bos|><|system_start|>tools:
-
story 3.76together and became good friends. And the ant never