Feature #382
Fires on “ friend” (not labeled yet: read the examples)
- Fires on
- 0.84% of tokens
- Mostly in
- story
- Layer
- after block 4
Where it fires
- story 72%
- chat 14%
- agent 10%
- python 3%
- worksheet 1%
Makes the model say
<|tool_call_end|> friend<|assistant_end|>." friends<|arg|> sorrydef
Makes it not say
nder counts anaking kickedy thing very
Strongest examples
On text the dictionary never trained on. The underlined token is where it fires most.
-
story 2.0but he could not. He was very sad. Then, a kind cat named Luna came to
-
story 1.96wanted to know what was inside.⏎Foxy called his friend, Birdy the
-
story 1.94sad, but it kept driving.⏎The dirty train met a friend, a big rain
-
story 1.93hurt and sad that he could not climb the chair.⏎Tim's mom saw him and said
-
story 1.93was so happy, and he shared the new ball with his new dog friend.<|bos|># requirements
-
story 1.92! A naughty dog had eaten it. Lily was sad. Her mom bought her another ice
-
story 1.91because he made a big mess.⏎Tom's friend, a small dog named Sam, came
-
story 1.9met a cute bunny. The bunny said
-
story 1.89s, plates, and yummy snacks. She invited her best friend, Timmy, to
-
story 1.87it was too hard for him.⏎Tom's friend, a little girl named Lucy, came
-
story 1.86jar, he found yummy cookies inside. He shared the cookies with his friend, Sam. Sam
-
story 1.86Bob was sad and did not know what to do. He asked his friend, the squirrel,