Feature #1086
Fires on “ st” (not labeled yet: read the examples)
- Fires on
- 1.49% of tokens
- Mostly in
- agent
- Layer
- after block 4
Where it fires
- agent 40%
- worksheet 22%
- python 14%
- chat 13%
- story 11%
Makes the model say
function<|assistant_end|>script wayfegevowelstail
Makes it not say
"`.#(- example inside⏎ ze
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.01, "Help! Help! I am scared!" Ben runs down the stairs. He sees
-
story 2.96ic door and check out what was inside. He started to climb the stairs and he noticed
-
chat 2.94. The princess was curious, so she ran up the stairs to invest
-
story 2.93around the house, in and out of rooms, up and down stairs, and all
-
chat 2.92and over again. The princess was curious, so she ran up the stairs to invest
-
story 2.9never find me there." She opens the door and goes down the stairs.⏎The
-
story 2.88yellow slide. It looked so inviting, so Molly ran up the stairs and zoomed
-
story 2.82The train's job was to attach carts and take them to the station. One day
-
story 2.68up and up. She was curious. She wanted to go up the stairs. She left
-
python 2.67sum_list` should add
-
story 2.65doll. She was happy. She was brave. She had climbed the stairs.<|bos|>Once
-
story 2.57shouts.⏎But Ben does not listen. He runs down the stairs and out of