Feature #1403
Fires on “3” (not labeled yet: read the examples)
- Fires on
- 0.39% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 70%
- chat 20%
- agent 6%
- story 3%
- python 2%
Makes the model say
times apples OneOne about doubleog high
Makes it not say
5hebow saw happily<|assistant_end|>ed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 4.899: 8+9=1
-
worksheet 4.845: 8+5=1
-
worksheet 4.817: 3+7=1
-
worksheet 4.817: 3+7=1
-
worksheet 4.856: 9+6=1
-
worksheet 4.6820: 8+0=8
-
chat 4.668: 3+8=1
-
worksheet 4.6565: 3+5+1
-
worksheet 4.635906: 7+6=1
-
worksheet 4.575: 2+5=7
-
worksheet 4.560: 9+0=9
-
worksheet 4.560: 9+0=9