Feature #52
Fires on “ went” (not labeled yet: read the examples)
- Fires on
- 2.05% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 84%
- chat 10%
- agent 4%
- python 2%
- worksheet 1%
Makes the model say
up to many andus allitems red
Makes it not say
.? trained<|assistant_end|> anymore hurt<|user_end|>")
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
chat 3.45was very hungry. He wanted to get some food. He went to the shop with
-
chat 3.33. Tim was very hungry. He wanted to get some food. He went to the shop with
-
story 3.31play his drawing. He wanted to show it to everyone. He went outside to put his
-
story 3.14girl named Sally wanted to make a special dish for her mom. She went to the kitchen and
-
story 3.1they decided to make a calendar for their mom. They used a big sheet
-
story 3.09. They wanted to see what was inside the boxes.⏎They climbed on the truck and
-
story 3.08adventure to find the best spot to bake her treats.⏎She walked and walked until she
-
story 3.03in his room. He was looking for his toy car. He looked under his bed, but
-
story 3.02adventure, so he decided to go exploring in the woods.⏎He walked for a long time
-
chat 2.97wanted to find a way to make the wheel feel better.⏎They went to a kind man
-
story 2.91it look nice, so she decided to polish it. Sue got a cloth and started to
-
story 2.9find his toy car. So, he started a search. He looked in the toy box,