Feature #2466
Fires on “5” (not labeled yet: read the examples)
- Fires on
- 1.12% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 84%
- chat 13%
- agent 2%
- story 1%
- python 0%
Makes the model say
+lessmer0uallyre resultra
Makes it not say
himself why let<|assistant_end|> !" yourself herself
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 7.162 = 550.⏎5749 + 32
-
worksheet 7.1329 = 84.⏎4325 + 27
-
worksheet 6.976 = 1560.⏎5577+522
-
worksheet 6.96547 = 9200.⏎7455 plus 8 equals
-
worksheet 6.864417.⏎The sum of 7253 and 21
-
worksheet 6.829 + 10 = 79.⏎68641 + 3
-
worksheet 6.76632 = 6603.⏎8930 + 89
-
worksheet 6.672988.⏎The sum of 9586 and 32
-
worksheet 6.59272 = 360.⏎6263 plus 7 equals
-
worksheet 6.514 + 8 = 12.⏎7311 + 92
-
worksheet 6.4839 = 17385.⏎83466 plus 4
-
worksheet 6.47365 = 7375.⏎6865 + 88