Feature #2722
Fires on “7” (not labeled yet: read the examples)
- Fires on
- 2.94% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 87%
- chat 6%
- agent 5%
- python 1%
- story 1%
Makes the model say
played play wave hurt playing67
Makes it not say
gerven applesOnce<|user_start|> example something
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.73+1=7, 70071⏎So 65675
-
worksheet 3.71+0=7, 78477⏎So 77705
-
worksheet 3.71, 78031⏎So 36 + 77995
-
worksheet 3.68=10, 106333⏎So 79153
-
worksheet 3.667, 77853⏎So 0 + 77853
-
worksheet 3.66+2=9, 93850⏎So 73213
-
worksheet 3.65+0=3, 36013⏎So 35944
-
worksheet 3.647, 75673⏎So 0 + 75673
-
worksheet 3.63+1=6, 64871⏎So 56994
-
worksheet 3.63=14, 140812⏎So 67718
-
worksheet 3.61+1=3, 30111⏎So 29964
-
worksheet 3.59+1=9, 98162⏎So 89798