Feature #2192
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 63%
- chat 30%
- story 4%
- agent 2%
- python 1%
Makes the model say
you# hel7, smiled we
Makes it not say
higher =elyvaluesmallestours closerx
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 8.21066+3886: 6+6=12, 2
-
worksheet 8.2⏎376+756: 6+6=12, 2
-
chat 8.09<|think_start|>516+056: 6+6=12, 2
-
worksheet 8.06006+3636: 6+6=12, 2
-
chat 8.04<|user_end|><|assistant_start|><|think_start|>06+76: 6+6=12, 2
-
worksheet 8.0⏎366+256: 6+6=12, 2
-
worksheet 7.8946+67726: 6+6=12, 2
-
worksheet 7.7146+88026: 6+6=12, 2
-
worksheet 7.719, 9⏎6+6: 6+6=12, 1
-
worksheet 7.626?⏎56+56: 6+6=12, 2
-
worksheet 7.5666+76396: 6+6=12, 2
-
worksheet 7.47667+5825: 7+5=12, 2