Feature #188
Fires on “+” (not labeled yet: read the examples)
- Fires on
- 0.65% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 59%
- chat 28%
- agent 6%
- story 4%
- python 3%
Makes the model say
= moreNA wouldnidesuck2
Makes it not say
and heard getso > curious$ has
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 7.319+4487
-
worksheet 7.319+4=13
-
worksheet 7.319+0+1=
-
chat 7.319+7=16
-
chat 7.319+8=17
-
worksheet 7.319+4=13
-
worksheet 7.22+1=8,
-
worksheet 7.22+6=12
-
chat 7.22+002:
-
worksheet 7.22+8: 4
-
worksheet 7.22+0=8,
-
chat 7.22+4=4,