Feature #2955
Fires on “3” (not labeled yet: read the examples)
- Fires on
- 1.11% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 72%
- chat 23%
- python 3%
- agent 1%
- story 0%
Makes the model say
anyoneun all blew average language mouth tried
Makes it not say
ugg ideaanceight huggedl])` hug
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.31, 74⏎053+448: 3+8+1
-
worksheet 4.246, 6⏎098+854: 8+4+1
-
worksheet 4.228, 864⏎61+94: 1+4+1
-
worksheet 4.21, 62⏎897+083: 7+3+1
-
worksheet 4.19, 13⏎988+060: 8+0+1
-
worksheet 4.191⏎0003+7799: 3+9+1
-
worksheet 4.121⏎2996+0014: 6+4+1
-
worksheet 4.11, 71⏎656+043: 6+3+1
-
worksheet 4.11, 58⏎178+695: 8+5+1
-
worksheet 4.13, 3⏎093+174: 3+4+1
-
worksheet 4.15⏎0349+6225: 9+5+1
-
worksheet 4.16⏎5203+4826: 3+6+1