Feature #2271
Fires on “ hurt” (not labeled yet: read the examples)
- Fires on
- 0.43% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 78%
- chat 10%
- agent 8%
- python 3%
- worksheet 1%
Makes the model say
them started very be eat picked wanted into
Makes it not say
_*.]`newreverseflip goodbye<|assistant_end|> best
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 3.04should never touch knives. They are very dangerous. They can hurt you badly."
-
story 2.98said, "Be careful with that fork, Lily. It can hurt you." He wanted
-
story 2.89. That's a razor. It's sharp and it can hurt you. Give
-
story 2.67it was a jellyfish. They said jellyfish can sting and hurt. They said jelly
-
story 2.59end, Lily and her brother learned to be careful with things that can hurt them. They knew
-
story 2.56⏎Lily was frightened. She knew guns were bad. They could hurt people. She wanted
-
story 2.35toys, but his favorite toy was a flexible snake. The snake could bend and twist in
-
story 2.31, that envelope is dangerous. We must not open it. It can hurt us."⏎Lily
-
story 2.31said they should never touch guns again. She said guns could hurt them and others.
-
story 2.26and said, "Tim, don't touch that bug. It might hurt you." But Tim
-
story 2.19Ben and Lily were playing with their toys in the living room. Ben had a big truck and
-
story 2.17play outside. They found some amazing sunglasses on the ground. The cat wanted to wear the sunglasses