Feature #3328
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 1.26% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 71%
- chat 25%
- agent 2%
- python 1%
- story 1%
Makes the model say
_")+ tests!")v)=
Makes it not say
itedBut Butes This from thoughake
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 7.016, 6⏎00+63: 0+3+1=4
-
worksheet 6.885, 5⏎43+31: 3+1+1=5
-
worksheet 6.85, 53⏎22+66: 2+6+1=9
-
worksheet 6.83516⏎00+48: 0+8+1=9
-
worksheet 6.82528⏎00+38: 0+8+1=9
-
worksheet 6.77418⏎00+38: 0+8+1=9
-
worksheet 6.745, 5⏎33+52: 3+2+1=6
-
worksheet 6.7311⏎299+001: 9+1+1=1
-
worksheet 6.725, 5⏎10+81: 0+1+1=2
-
worksheet 6.7102⏎086+343: 6+3+1=1
-
worksheet 6.71, 40⏎12+00: 2+0+1=3
-
worksheet 6.7, 84⏎00+11: 0+1+1=2