Feature #2602
Fires on “6” (not labeled yet: read the examples)
- Fires on
- 0.70% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 85%
- chat 7%
- agent 5%
- story 2%
- python 1%
Makes the model say
<|think_end|>d⏎ difference woods listnumberize
Makes it not say
he *nase new bi It
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 5.375+7+1=13, 132509⏎So
-
worksheet 4.98: 0+5=5, 55679⏎So 2
-
worksheet 4.984+2+1=7, 74763⏎So 4
-
worksheet 4.916+8=14, 141697⏎So 6
-
worksheet 4.80: 4+0=4, 41606⏎So 4
-
worksheet 4.786+7=13, 138105⏎So 6
-
worksheet 4.78: 0+8=8, 86164⏎So 8
-
worksheet 4.74: 0+3=3, 37726⏎So 8
-
worksheet 4.6: 0+6+1=7, 7666⏎0+3
-
worksheet 4.574+2+1=7, 72876⏎So 4
-
worksheet 4.56+5+1=11, 114928⏎So 5
-
worksheet 4.47+5+1=12, 121682⏎So 6