Feature #2413
Fires on “+” (not labeled yet: read the examples)
- Fires on
- 0.39% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 68%
- chat 20%
- agent 6%
- python 4%
- story 2%
Makes the model say
ladyoddv<|assistant_end|> bothving elementde
Makes it not say
languageel!")obit next:ut
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.51, 1⏎60+05: 0+5+1=6,
-
worksheet 2.52, 23⏎9+0: 9+0+1=10
-
chat 2.52, 20⏎4+0: 4+0+1=5,
-
chat 2.5, 1⏎14+00: 4+0+1=5,
-
chat 2.48, 3⏎37+00: 7+0+1=8,
-
worksheet 2.486, 60⏎0+2: 0+2+1=3,
-
worksheet 2.472, 26⏎2+0: 2+0+1=3,
-
chat 2.476, 65⏎9+2: 9+2+1=12
-
chat 2.463, 30⏎2+0: 2+0+1=3,
-
chat 2.4610, 0⏎5+0: 5+0+1=6,
-
worksheet 2.4523⏎36+02: 6+2+1=9,
-
worksheet 2.4413⏎00+70: 0+0+1=1,