Feature #3593
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 1.48% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 61%
- chat 25%
- agent 7%
- python 4%
- story 4%
Makes the model say
]) + `[True])`=()` no
Makes it not say
ish9 houseot putlou But
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.86=8, 84⏎5+0: 5+0=5,
-
chat 3.83=5, 52⏎5+0: 5+0=5,
-
chat 3.822⏎5+0: 5+0=5,
-
worksheet 3.754, 480⏎5+0: 5+0=5,
-
worksheet 3.75=8, 83⏎6+0: 6+0=6,
-
worksheet 3.73⏎4+0: 4+0=4,
-
worksheet 3.72, 5547⏎6+0: 6+0=6,
-
chat 3.71=6, 61⏎6+0: 6+0=6,
-
worksheet 3.697⏎000+600: 0+0=0,
-
worksheet 3.68⏎5+5: 5+5=10,
-
worksheet 3.67=8, 88⏎5+0: 5+0=5,
-
worksheet 3.67=7, 71⏎3+0: 3+0=3,