Feature #2607
Fires on “3” (not labeled yet: read the examples)
- Fires on
- 0.70% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 80%
- agent 8%
- worksheet 6%
- python 4%
- story 2%
Makes the model say
8025374ant
Makes it not say
(A ( upon?"$ce lived
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.55+1: 3+1=4, 48<|think_end|>48<|assistant_end|><|bos|><|system_start|>
-
chat 2.51+2: 1+2=3, 38<|think_end|>38<|assistant_end|><|bos|>F
-
chat 2.50: 3+0=3, 323<|think_end|>323<|assistant_end|><|bos|>
-
chat 2.490: 3+0=3, 323<|think_end|>323<|assistant_end|><|bos|>
-
chat 2.49+0: 2+0=2, 27<|think_end|>27<|assistant_end|><|bos|>If
-
chat 2.49+1: 0+1=1, 13<|think_end|>13<|assistant_end|><|bos|>Once
-
chat 2.49+0: 4+0=4, 43<|think_end|>43<|assistant_end|><|bos|><|system_start|>
-
chat 2.47=9, 976<|think_end|>976<|assistant_end|><|bos|><|system_start|>
-
chat 2.47: 0+3+1=4, 47<|think_end|>47<|assistant_end|><|bos|>what
-
chat 2.467: 0+7=7, 790<|think_end|>790<|assistant_end|><|bos|>
-
chat 2.450: 3+0=3, 318<|think_end|>318<|assistant_end|><|bos|>
-
chat 2.450=2, 297<|think_end|>297<|assistant_end|><|bos|><|system_start|>