Feature #637
Fires on “ unexpected” (not labeled yet: read the examples)
- Fires on
- 0.36% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 64%
- story 26%
- worksheet 4%
- chat 4%
- python 1%
Makes the model say
out happenedpositive happen sorryodd<|assistant_end|> work
Makes it not say
python7Once import sunONT]`.
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.64fill the hole with sand.⏎But then, something unexpected happened. As Tim
-
story 3.64with the basketball. They were very happy. But then, something unexpected happened. The basket
-
chat 3.54to race and Cari was winning. But then, something unexpected happened. Car
-
story 3.54baby nice and clean.⏎But, then something unexpected happened! The baby
-
story 3.53under a tree. Sam was so happy!⏎But then, something unexpected happened. The big
-
story 3.52together. They were very happy.⏎But then, something unexpected happened. A big
-
story 3.52anymore. She liked the toy a lot.⏎But then, something unexpected happened. The scary
-
chat 3.51the tasty cake. Tim was so happy!⏎But then, something unexpected happened. The cat
-
story 3.47She went fast and had a lot of fun. But then, something unexpected happened. The k
-
story 3.46Tim for his help.⏎But then, something unexpected happened. The wind
-
story 3.46are!" Lucy said. She was very happy. But then, something unexpected happened. A small
-
story 3.44friend. They talked and played all day.⏎But then, something unexpected happened. The flower