Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #3741

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

Fires on
0.33% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 70%
  • story 22%
  • chat 5%
  • worksheet 2%
  • python 1%

Makes the model say

? friend new turns<|tool_call_end|> while share!"

Makes it not say

faces never upperenties question had always

Strongest examples

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

  1. story 3.74
    , but he knew she was really worried. He wanted to help, so he said, "
  2. story 3.69
    something very important on his desk. He wanted to protect it, so he sat very still
  3. story 3.52
    small cat stuck in the mud. Spot wanted to help the cat, so he tried to pull
  4. story 3.51
    The frog looked sick and tired. Tom wanted to help the frog, so he took it out
  5. story 3.47
    . The bug was slow too. Amy wanted to help the bug, so she put it on
  6. story 3.46
    was hurt and couldn't fly. Tim wanted to help the bird, so he picked it up
  7. story 3.44
    She had a clean towel with her. Sue wanted to help Spot, so she tried to give
  8. story 3.43
    I lost them when I was playing." Spot wanted to help Tim, so they both started to
  9. story 3.42
    could see it was mail for someone. Lilly wanted to help, so she took the letter
  10. story 3.4
    could not find his way home. Elly wanted to help Billy, so she said, "
  11. story 3.39
    that it was hard to see its color. Tim wanted to help, so he decided to wash
  12. story 3.39
    bear and I can't find him." Jim wanted to help Sue, so they looked for the