Feature #3217
Fires on “=” (not labeled yet: read the examples)
- Fires on
- 0.48% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 54%
- agent 23%
- chat 21%
- story 1%
- python 1%
Makes the model say
)`. differenceP= wind of thingsYes
Makes it not say
0 sumhere<|assistant_end|>riteOneed
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.17: 7+9+1=17,
-
worksheet 4.1356⏎6+9: 6+9+1=16,
-
worksheet 4.1179+0836: 9+6+1=16,
-
chat 4.0891⏎9+6: 9+6+1=16,
-
worksheet 4.031⏎09+49: 9+9+1=19,
-
chat 4.02, 1⏎6+9: 6+9+1=16,
-
chat 4.023⏎09+66: 9+6+1=16,
-
worksheet 3.9816+9769: 6+9+1=16,
-
chat 3.9791⏎9+6: 9+6+1=16,
-
worksheet 3.975⏎19+39: 9+9+1=19,
-
worksheet 3.96385⏎7+9: 7+9+1=17,
-
chat 3.95, 1⏎6+9: 6+9+1=16,