Feature #2927
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 1.22% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 80%
- chat 13%
- agent 6%
- python 1%
- story 1%
Makes the model say
⏎,!"." laughed?"<|think_end|>name
Makes it not say
k inv reallyprint tablecalculatorif
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.972: 0+2=2, 2973⏎0+7
-
worksheet 3.890: 4+0=4, 41606⏎So
-
worksheet 3.8975: 0+5=5, 5673⏎0+7
-
worksheet 3.860: 9+0=9, 9622⏎So 9
-
worksheet 3.850: 9+0=9, 9262⏎So 9
-
worksheet 3.8365: 0+5=5, 5234⏎0+6
-
worksheet 3.79805: 0+5=5, 590⏎00+
-
worksheet 3.754: 0+4=4, 4702⏎So 5
-
worksheet 3.7500: 9+0=9, 9810⏎4+0
-
worksheet 3.7400: 6+0=6, 6264⏎6+0
-
worksheet 3.7281: 0+1=1, 1697⏎6+8
-
worksheet 3.7200: 1+0=1, 1854⏎5+0