Feature #1512
Fires on “ friends” (not labeled yet: read the examples)
- Fires on
- 0.21% of tokens
- Mostly in
- python
- Layer
- after block 4
Where it fires
- python 44%
- story 41%
- chat 8%
- agent 6%
- worksheet 1%
Makes the model say
m were whoal with is again best
Makes it not say
```]`. Finopx<|bos|>RThis
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 5.977 = 92725.<|bos|>Lily and Tom are friends. They like to
-
story 5.88<|think_end|>The answer is 78.<|assistant_end|><|bos|>Tom and Sam are friends. They like to
-
story 5.87790<|think_end|>790<|assistant_end|><|bos|>Lily and Ben are friends. They like to
-
story 5.817 = 91540<|bos|>Tom and Mia were best friends. They lived next
-
story 5.724, 9])⏎9⏎<|bos|>Tom and Ann are friends. They like to
-
story 5.6only tell short stories and add numbers.<|assistant_end|><|bos|>Tom and Sam were friends. They liked to
-
story 5.55_stats.py⏎all tests passed⏎<|bos|>Tom and Jen are friends. They like to
-
story 5.55answer is 1706.<|assistant_end|><|bos|>Tim and Sam are friends. They like to
-
story 5.48<|think_end|>The answer is 63.<|assistant_end|><|bos|>Anna and Ben are friends. They like to
-
story 5.47very happy to have each other as friends.<|bos|>Tom and Lily are friends. They like to
-
story 5.4795 is 2603.<|bos|>Anna and Ben are friends. They like to
-
story 5.46was happy to have a new friend.<|assistant_end|><|bos|>Tom and Mia are friends. They like to