Feature #363
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.64% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 58%
- chat 28%
- agent 8%
- story 5%
- python 1%
Makes the model say
opencodere lastleLookT workscript
Makes it not say
3pydef sad into small# to
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.36<|user_end|><|assistant_start|><|think_start|>9+1: 9+1=10,
-
chat 2.24+ 9?<|user_end|><|assistant_start|><|think_start|>9+9: 9+9=18,
-
chat 2.239 for me?<|user_end|><|assistant_start|><|think_start|>8+9: 8+9=17,
-
chat 2.17<|user_end|><|assistant_start|><|think_start|>7+1: 7+1=8,
-
chat 2.169⏎3+0: 3+0=3,
-
chat 2.074⏎37+01: 7+1+1=9
-
chat 2.052.<|user_end|><|assistant_start|><|think_start|>83+92: 3+2=5,
-
chat 2.048, 8⏎3+1: 3+1=4,
-
worksheet 2.01, 41⏎09+49: 9+9+1=1
-
chat 2.01+ 5?<|user_end|><|assistant_start|><|think_start|>8+5: 8+5=13,
-
chat 2.01+8: 8+8=16,
-
chat 2.0, 6⏎88+81: 8+1+1=1