Feature #2587
Fires on “ +” (not labeled yet: read the examples)
- Fires on
- 0.37% of tokens
- Mostly in
- worksheet
- Layer
- after block 4
Where it fires
- worksheet 80%
- chat 14%
- story 3%
- agent 2%
- python 0%
Makes the model say
and butWprint even printselyNo
Makes it not say
Sara<|assistant_end|> t0tomean * seen
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
worksheet 4.6510 = 79.⏎68641 + 356
-
story 4.58the park and in the house. One day, they decided to play hide and seek. Lily
-
worksheet 4.5455 is 100.⏎44368 + 983
-
story 4.42in the park with their gray dog, Max. They like to play hide and seek with Max
-
story 4.34and they started to play an interesting game.⏎"Let's play hide and seek," one
-
worksheet 4.33= 72876.⏎89192 + 384
-
worksheet 4.29= 92304.⏎45564 + 848
-
worksheet 4.268 = 43109.⏎5167 + 15 =
-
worksheet 4.2465 = 7375.⏎6865 + 887
-
worksheet 4.2494 = 9593⏎76953 + 21 =
-
worksheet 4.213 = 1255.⏎78140 + 640
-
worksheet 4.28 equals 7463.⏎40797 + 6 =