Feature #372
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.40% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 43%
- chat 28%
- agent 22%
- story 4%
- python 2%
Makes the model say
duate noise[- //relK hugs
Makes it not say
He therest of and that he out
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.63<|assistant_start|><|think_start|>743+422: 3+2=5, 5
-
chat 7.57<|assistant_start|><|think_start|>343+212: 3+2=5, 5
-
chat 7.37?<|user_end|><|assistant_start|><|think_start|>22+03: 2+3=5, 5
-
chat 7.33?<|user_end|><|assistant_start|><|think_start|>93+02: 3+2=5, 5
-
chat 7.29<|assistant_start|><|think_start|>722+793: 2+3=5, 5
-
worksheet 7.28242+04573: 2+3=5, 5
-
chat 7.18the result.<|user_end|><|assistant_start|><|think_start|>3+2: 3+2=5, 5
-
chat 7.15<|assistant_start|><|think_start|>393+052: 3+2=5, 5
-
chat 7.14?<|user_end|><|assistant_start|><|think_start|>52+03: 2+3=5, 5
-
chat 7.13<|assistant_start|><|think_start|>743+422: 3+2=5, 5
-
chat 7.1<|assistant_start|><|think_start|>852+443: 2+3=5, 5
-
chat 7.09.<|user_end|><|assistant_start|><|think_start|>83+92: 3+2=5, 5