Feature #256
Fires on “0” (not labeled yet: read the examples)
- Fires on
- 1.40% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 74%
- chat 22%
- agent 3%
- story 1%
- python 1%
Makes the model say
+[- help even yellow always fun knew
Makes it not say
:en tests," treesage greetingure
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 7.71⏎9280+0560: 0+0+1=1
-
worksheet 7.595⏎540+000: 0+0+1=1
-
worksheet 7.496⏎832+000: 2+0+1=3
-
worksheet 7.45⏎9217+0000: 7+0+1=8
-
worksheet 7.41⏎2766+0021: 6+1+1=8
-
worksheet 7.366⏎090+560: 0+0+1=1
-
worksheet 7.346⏎814+000: 4+0+1=5
-
worksheet 7.336⏎5203+4826: 3+6+1
-
worksheet 7.3355⏎933+000: 3+0+1=4
-
worksheet 7.3158⏎895+070: 5+0+1=6
-
worksheet 7.2411⏎957+634: 7+4+1=1
-
worksheet 7.215⏎238+000: 8+0+