Feature #1026
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.31% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 46%
- chat 25%
- agent 21%
- story 4%
- python 4%
Makes the model say
my nowed sorryi. anymore,"
Makes it not say
tableOnce D many daddy It#One
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 2.72: 6+2=8, 8
-
worksheet 2.56: 2+6=8, 8
-
chat 2.49=3, 3⏎6+2: 6+2=8, 8
-
worksheet 2.46does Max have now?⏎6+2: 6+2=8, 8
-
worksheet 2.39: 2+6=8, 8
-
chat 2.38=3, 3⏎6+2: 6+2=8, 8
-
chat 2.31have now?<|user_end|><|assistant_start|><|think_start|>2+6: 2+6=8, 8
-
chat 2.22=5, 5⏎2+6: 2+6=8, 8
-
worksheet 2.18, 488⏎2+6: 2+6=8, 8
-
worksheet 2.18086+72522: 6+2=8, 8
-
worksheet 2.17?⏎052+306: 2+6=8, 8
-
worksheet 2.16, 953⏎2+6: 2+6=8, 8