Feature #1642
Fires on “4” (not labeled yet: read the examples)
- Fires on
- 0.71% of tokens
- Mostly in
- chat
- Layer
- after block 4
Where it fires
- chat 47%
- worksheet 26%
- agent 21%
- story 5%
- python 1%
Makes the model say
They( is true plus they can app
Makes it not say
."final againbye shouted Butly case
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 3.16Now Mia has 8 coins.⏎7 plus 3964 is 39
-
worksheet 3.01had fun at the zoo.<|assistant_end|><|bos|>7 plus 44311 equals 4
-
worksheet 2.97= 3676.⏎1 plus 96223 is 9
-
worksheet 2.9462 = 63.⏎2 plus 76146 equals 7
-
worksheet 2.9113289.⏎6 plus 5260 is 52
-
worksheet 2.84144022.⏎3 plus 9976 is 99
-
worksheet 2.74.<|assistant_end|><|bos|>How much is 1 + 41532?⏎00
-
worksheet 2.728.⏎6 + 4544 = 45
-
worksheet 2.69together and had lots of fun.<|bos|>7 + 3000 = ?⏎0
-
worksheet 2.691 is 108.⏎8 + 7539 = 75
-
worksheet 2.699⏎So 7 + 6242 = 62
-
worksheet 2.6737 = 172⏎9 plus 8370 is 83