Feature #951
Fires on “9” (not labeled yet: read the examples)
- Fires on
- 0.54% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 38%
- chat 32%
- python 12%
- story 10%
- worksheet 9%
Makes the model say
to's forasing things fish home
Makes it not say
[- smallest$(- Python twins=**/*.ose
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 1.93<|assistant_start|>The answer is 9.<|assistant_end|><|user_start|>And if I add 54?<|user_end|><|assistant_start|>
-
chat 1.88The answer is 59.<|assistant_end|><|user_start|>And if I add 19?<|user_end|><|assistant_start|>
-
chat 1.76<|assistant_start|>The answer is 0.<|assistant_end|><|user_start|>And if I add 37?<|user_end|><|assistant_start|>
-
chat 1.76.<|assistant_end|><|user_start|>And if I add 5?<|user_end|><|assistant_start|><|think_start|>
-
chat 1.73The answer is 38.<|assistant_end|><|user_start|>And if I add 98?<|user_end|><|assistant_start|>
-
chat 1.73The answer is 50.<|assistant_end|><|user_start|>And if I add 46?<|user_end|><|assistant_start|>
-
chat 1.72The answer is 41.<|assistant_end|><|user_start|>And if I add 16?<|user_end|><|assistant_start|>
-
chat 1.66answer is 7.<|assistant_end|><|user_start|>What if you add 95 more?<|user_end|>
-
chat 1.63answer is 610.<|assistant_end|><|user_start|>What if you add 0 more?<|user_end|><|assistant_start|>
-
chat 1.62The answer is 49.<|assistant_end|><|user_start|>And if I add 92?<|user_end|><|assistant_start|>
-
chat 1.62The answer is 15.<|assistant_end|><|user_start|>What if you add 0 more?<|user_end|><|assistant_start|>
-
chat 1.61The answer is 41.<|assistant_end|><|user_start|>And if I add 8?<|user_end|><|assistant_start|>The