Feature #19
Fires on “ equals” (not labeled yet: read the examples)
- Fires on
- 0.83% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 66%
- chat 18%
- story 7%
- agent 5%
- python 5%
Makes the model say
=`.alse"))"`.:`.ace [
Makes it not say
himselfake floorMoment woke woulds
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.2636118 balls.⏎82 plus 66 equals 148
-
worksheet 4.2412546.⏎5 plus 783 equals 788
-
worksheet 4.189 is 5182.⏎8 plus 25 equals 33.
-
worksheet 4.080 = 274.⏎51 plus 91 equals 142
-
worksheet 4.088 + 5 = 13⏎99 plus 9 equals 108
-
worksheet 4.0169389.⏎8 plus 146 equals 154
-
worksheet 3.9949 = 155.⏎98 plus 6 equals 104
-
worksheet 3.9686 = 83882⏎7 plus 5 equals 12.
-
worksheet 3.95= 71402⏎70 plus 60 equals 130
-
worksheet 3.93102869.⏎8 plus 85 equals 93.
-
worksheet 3.938 = 29473.⏎7 plus 7 equals 14.
-
worksheet 3.92Now Lucy has 61 coins.⏎5 plus 26 equals 31.