Feature #772
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.27% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 57%
- chat 36%
- python 3%
- agent 3%
- story 2%
Makes the model say
or4 cars ownfrom>>> We
Makes it not say
ite charactersemember])) aboutit runningation
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 8.57⏎8+7: 8+7=15, 1592
-
chat 8.47+8: 8+8=16, 167<|think_end|>
-
chat 8.33<|assistant_start|><|think_start|>8+9: 8+9=17, 17<|think_end|>The
-
worksheet 8.287+8: 7+8=15, 1579
-
worksheet 8.184⏎7+8: 7+8=15, 1581
-
chat 8.16<|assistant_start|><|think_start|>9+1: 9+1=10, 10<|think_end|>The
-
chat 8.13<|assistant_start|><|think_start|>9+9: 9+9=18, 18<|think_end|>The
-
chat 8.06<|assistant_start|><|think_start|>9+8: 9+8=17, 17<|think_end|>The
-
worksheet 8.030⏎9+8: 9+8=17, 1758
-
worksheet 8.033⏎9+7: 9+7=16, 1632
-
chat 8.01<|assistant_start|><|think_start|>8+9: 8+9=17, 17<|think_end|>The
-
worksheet 7.94+8: 4+8=12, 12⏎So