Feature #906
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.35% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 71%
- python 15%
- agent 7%
- chat 5%
- story 1%
Makes the model say
Created{ + function "$⏎ def()`
Makes it not say
9 animals it<|bos|>One!Falseel
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.71066+3886: 6+6=12, 2
-
worksheet 3.66⏎376+756: 6+6=12, 2
-
chat 3.49<|user_end|><|assistant_start|><|think_start|>06+76: 6+6=12, 2
-
chat 3.34<|think_start|>516+056: 6+6=12, 2
-
worksheet 3.3346+67726: 6+6=12, 2
-
worksheet 3.3226+27: 6+7=13, 3
-
worksheet 3.27006+3636: 6+6=12, 2
-
worksheet 3.2546+88026: 6+6=12, 2
-
worksheet 3.24⏎366+256: 6+6=12, 2
-
chat 3.22<|think_start|>295+177: 5+7=12, 2
-
worksheet 3.22667+5825: 7+5=12, 2
-
chat 3.22846+767: 6+7=13, 3