Feature #1758
Fires on “⏎” (not labeled yet: read the examples)
- Fires on
- 0.67% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 67%
- chat 16%
- python 14%
- agent 2%
- story 1%
Makes the model say
⏎demoflip reverreverses called window
Makes it not say
able I):i's story main acc
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.36: 7+0+1=8, 801⏎16+0
-
worksheet 3.270: 6+0+1=7, 76⏎70+1
-
worksheet 3.25: 2+0+1=3, 340⏎1+0:
-
worksheet 3.25: 5+0+1=6, 608⏎4+0:
-
worksheet 3.236: 2+6+1=9, 97⏎98+0
-
worksheet 3.2200: 0+0=0, 079⏎97+0
-
chat 3.2+00: 1+0=1, 11⏎4+0:
-
worksheet 3.2: 7+0+1=8, 810⏎87+0
-
worksheet 3.1959: 0+9=9, 913⏎0+5:
-
worksheet 3.18: 2+0+1=3, 300⏎97+0
-
worksheet 3.1700: 9+0=9, 983⏎58+0
-
worksheet 3.1700: 0+0=0, 034⏎00+8