Feature #15
Fires on “:” (not labeled yet: read the examples)
- Fires on
- 0.75% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 76%
- chat 22%
- story 1%
- python 1%
- agent 0%
Makes the model say
python python,"<|assistant_start|>Once: promised put
Makes it not say
3 x24 like gone came thing
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 1.714.⏎01444+25174: 4+4
-
worksheet 1.6now?⏎64895+07613: 5+3
-
worksheet 1.64.⏎62705+15834: 5+4
-
worksheet 1.586.⏎27161+34136: 1+6
-
worksheet 1.584.⏎96122+47914: 2+4
-
worksheet 1.58me?⏎61178+43771: 8+1
-
worksheet 1.5843⏎02635+61143: 5+3
-
worksheet 1.581 for me?⏎0002+1331: 2+1
-
worksheet 1.566?⏎51523+51346: 3+6
-
worksheet 1.56Tim have now?⏎4326+2164: 6+4
-
worksheet 1.542?⏎00001+41532: 1+2
-
worksheet 1.5390.⏎Compute 5+3⏎5+3: 5+3