Feature #7
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 1.50% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 74%
- chat 23%
- agent 1%
- python 1%
- story 0%
Makes the model say
):() yourely nums my4Amy
Makes it not say
⏎. ItIt subtract un andw
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.446⏎0+4: 0+4=4, 4526⏎
-
worksheet 3.366+9: 6+9=15, 156864
-
worksheet 3.326: 8+6+1=15, 157258
-
worksheet 3.327+8: 7+8=15, 158144
-
worksheet 3.318⏎0+3: 0+3=3, 35158
-
worksheet 3.33⏎0+7: 0+7=7, 72973
-
worksheet 3.36+8: 6+8=14, 141697
-
worksheet 3.298⏎0+3: 0+3=3, 3318⏎
-
chat 3.284⏎5+0: 5+0=5, 584<|think_end|>5
-
worksheet 3.279⏎0+3: 0+3=3, 3999⏎
-
worksheet 3.275: 9+5+1=15, 15465⏎
-
worksheet 3.272⏎0+3: 0+3=3, 3092⏎