Feature #3731
Fires on “2” (not labeled yet: read the examples)
- Fires on
- 0.38% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 68%
- chat 13%
- story 11%
- agent 7%
- python 1%
Makes the model say
Once+ but + prints)` suddenly saw
Makes it not say
right moralfulen flip betterared
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.23So 48 + 23 = 71.⏎3+498
-
worksheet 3.23⏎So 1 + 62 = 63.⏎2 plus 76
-
worksheet 3.182⏎So 4 + 8 = 12.⏎7311 +
-
worksheet 3.1129 = 84.⏎4325 +
-
worksheet 3.1144 + 59 = 103.⏎20223
-
worksheet 3.114882 = 71441.⏎24232
-
worksheet 3.1128 + 21 = 149.⏎2504 plus
-
worksheet 3.1So 18 + 70 = 88.⏎2+155
-
agent 3.09= 0⏎13: for
-
worksheet 3.06So 13 + 12 = 25.⏎2365 plus
-
worksheet 3.051⏎So 2 + 9 = 11.⏎23 plus 9
-
worksheet 3.0441 + 72 = 113.⏎42+5=