Feature #1982
Fires on “ a” (not labeled yet: read the examples)
- Fires on
- 0.54% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 62%
- story 15%
- chat 9%
- worksheet 8%
- python 5%
Makes the model say
ands safe knew-r,u
Makes it not say
answersum sum<|think_end|> prelude Just sitonfig
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
python 2.19count_vowels(s
-
story 2.08, and that means slowly too."⏎Sally thought about this for a moment. "I
-
story 2.01of where I can find food."⏎The bird thought about this for a moment. Then he
-
story 1.96always remember the fun times you have with your friends. Recording these moments
-
story 1.94Bunny wanted to help Bouncy.⏎Bunny thought for a moment and then said
-
story 1.88snow with the kids, but I can't." The squirrel thought for a moment and said,
-
story 1.87or a picture," the voice said.⏎Tom and Anna thought for a moment. They wanted
-
story 1.84my steam is not strong."⏎The wise old tree thought for a moment and said,
-
chat 1.8<|system_start|>You are a helpful assistant.<|system_end|><|user_start|>Find the sum of 93 and
-
story 1.78he was walking, Tim met a friendly dog. The dog wagged its tail and barked. Tim
-
story 1.78it looks normal. Can you add something special?"⏎Lily thought for a moment and decided to
-
story 1.75help me learn the rhythm?" The frog thought for a moment and then replied