Skip to content
mini-lab
Features of prelude-2

Feature #2313

Fires on “ I” (not labeled yet: read the examples)

Fires on
0.31% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 39%
  • story 31%
  • agent 30%
  • worksheet 1%
  • python 1%

Makes the model say

[-et The justre wasutils me

Makes it not say

+ = longer plus?w x tell

Strongest examples

On text the dictionary never trained on. The underlined token is where it fires most.

  1. story 4.88
    bird eggs. They said, "We are sorry, big bird. We will leave your eggs
  2. story 4.8
    "We are sorry, mom. We did not think about Spot. We will share with him
  3. story 4.73
    ," they said. "We did not mean to hurt the veil. We just wanted to play
  4. chat 4.67
    zil?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
  5. story 4.65
    ⏎"I'm sorry, mom. I'm sorry, Lily. I was playing with my
  6. story 4.58
    I'm sorry. I didn't know it was her baby doll. I thought it was a
  7. story 4.58
    ," Tom says.⏎"We are sorry, Mom and Dad. We will not do it
  8. chat 4.56
    zil.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
  9. chat 4.54
    uliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
  10. chat 4.5
    ada?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
  11. chat 4.48
    uliet?<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short
  12. chat 4.42
    to me.<|user_end|><|assistant_start|>Sorry, I don't know about that. I can only tell short