Feature #1130
Fires on “ every” (not labeled yet: read the examples)
- Fires on
- 1.24% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 31%
- story 29%
- python 16%
- chat 14%
- worksheet 9%
Makes the model say
storyS Pythonitem import_ typoually
Makes it not say
shiny but full big behind could Can very
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.73⏎<|bos|><|system_start|>Start every answer with "Sure
-
python 3.51⏎A small Python project.
-
story 3.47baby monkey and they all played together every day.⏎The moral of the story is to always
-
chat 3.40], 8)⏎True⏎<|bos|><|system_start|>Start every answer with "Sure
-
chat 3.36the garden.<|bos|><|system_start|>Start every answer with "Sure
-
python 3.33⏎⏎>>> last_letter("
-
story 3.33ed to the garden.⏎The moral of this story is that when
-
python 3.3⏎⏎assert exclaim("banana")
-
python 3.3⏎⏎assert combine("robot",
-
story 3.3They both grew happier and wiser.⏎The moral of the story is: T
-
story 3.26to her friends, and they were happy too.⏎The moral of the story is to always
-
python 3.24⏎>>> lower_case("