Feature #1532
Fires on “ ” (not labeled yet: read the examples)
- Fires on
- 1.53% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 60%
- chat 20%
- agent 15%
- story 4%
- python 2%
Makes the model say
53,6): tests away2
Makes it not say
sunisu startedbsub wanted T
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 1.96⏎048+191: 8+1=
-
worksheet 1.958⏎450+009: 0+9+
-
worksheet 1.92⏎481+927: 1+7+
-
worksheet 1.91⏎756+094: 6+4=
-
worksheet 1.919⏎53+01: 3+1+
-
worksheet 1.893⏎441+009: 1+9=
-
worksheet 1.89⏎04+78: 4+8=
-
chat 1.84<|system_start|>Do not use the calculator
-
worksheet 1.82=8, 810⏎49+00: 9+0=
-
chat 1.82⏎0+9: 0+9=
-
chat 1.82+0=9, 9⏎54+00: 4+0=
-
worksheet 1.81+8=8, 8⏎79+00: 9+0=