Feature #5
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 1.16% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 67%
- chat 23%
- agent 6%
- python 4%
- story 1%
Makes the model say
+",how]) ` )")
Makes it not say
n friendanit printeda`.ome
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.910003: 0+3=3
-
chat 3.89031: 5+1=6
-
worksheet 3.8700053: 4+3=7
-
worksheet 3.8572503: 5+3=8
-
worksheet 3.821⏎0003+7799: 3+9+1
-
worksheet 3.7801: 5+1=6
-
worksheet 3.7780027: 7+7=1
-
worksheet 3.7630: 5+0=5
-
worksheet 3.725906: 7+6=1
-
worksheet 3.61⏎2996+0014: 6+4+1
-
worksheet 3.581⏎1679+0000: 9+0+1
-
worksheet 3.570⏎8776+0004: 6+4+1