Feature #3156
Fires on “1” (not labeled yet: read the examples)
- Fires on
- 0.69% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 64%
- chat 25%
- agent 7%
- story 2%
- python 2%
Makes the model say
<|arg|>ectpath= absolute ` laughed`.
Makes it not say
=, small cubed enough toy softf
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 2.3640⏎55+42: 5+2+1=8,
-
worksheet 2.33, 83⏎4+6: 4+6+1=11,
-
chat 2.298, 8⏎5+1: 5+1+1=7,
-
chat 2.25, 06⏎5+1: 5+1+1=7,
-
chat 2.248, 8⏎5+1: 5+1+1=7,
-
chat 2.238⏎5+1: 5+1+1=7,
-
worksheet 2.22733⏎5+2: 5+2+1=8,
-
worksheet 2.21941⏎0+4: 0+4+1=5,
-
worksheet 2.2164⏎61+94: 1+4+1=6,
-
worksheet 2.1921⏎75+92: 5+2+1=8,
-
worksheet 2.1708⏎45+00: 5+0+1=6,
-
chat 2.171+0: 1+0+1=2,