Feature #2625
Fires on “ friends” (not labeled yet: read the examples)
- Fires on
- 0.99% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 75%
- chat 10%
- agent 10%
- worksheet 3%
- python 2%
Makes the model say
. who." forever<|bos|> too learned liked
Makes it not say
schema Thennewsist thenhesly⏎
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 4.457 = 91540<|bos|>Tom and Mia were best friends. They lived next
-
story 4.21790<|think_end|>790<|assistant_end|><|bos|>Lily and Ben are friends. They like to
-
story 4.18They became friends and liked to play
-
story 4.147 = 92725.<|bos|>Lily and Tom are friends. They like to
-
story 4.1<|think_end|>The answer is 78.<|assistant_end|><|bos|>Tom and Sam are friends. They like to
-
story 4.09to tell stories and add numbers.<|assistant_end|><|bos|>Anna and Lily were best friends. They liked to
-
story 4.01_stats.py⏎all tests passed⏎<|bos|>Tom and Jen are friends. They like to
-
story 3.974, 9])⏎9⏎<|bos|>Tom and Ann are friends. They like to
-
story 3.95best eggs ever.⏎The end.<|bos|>Lily and Ben were best friends. They liked to
-
story 3.94=6, 6<|think_end|>6<|assistant_end|><|bos|>Lily and Tom were best friends. They liked to
-
story 3.87they had many fun adventures together.<|assistant_end|><|bos|>Ben and Sam were best friends. They liked to
-
story 3.85very happy to have each other as friends.<|bos|>Tom and Lily are friends. They like to