Feature #3408
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 0.63% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 78%
- chat 21%
- agent 1%
- story 0%
- python 0%
Makes the model say
lyared up<|think_start|> very floor inore
Makes it not say
):def_ they>>> also soon rain
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.11685⏎8965+4685: 5+5=
-
worksheet 4.05356⏎0184+4356: 4+6=
-
worksheet 4.0436?⏎0006+3636: 6+6=
-
worksheet 3.98866⏎6137+7866: 7+6=
-
worksheet 3.97019⏎4730+3019: 0+9=
-
worksheet 3.97402⏎8661+6402: 1+2=
-
worksheet 3.97063⏎0806+1063: 6+3=
-
worksheet 3.96493⏎0457+8493: 7+3=
-
worksheet 3.94, 0⏎4035+0815: 5+5+
-
worksheet 3.94, 3⏎8868+0353: 8+3+
-
worksheet 3.9465?⏎0007+4265: 7+5=
-
worksheet 3.9489=⏎0003+4989: 3+9=