Skip to content
mini-lab
Features of prelude-2

Feature #3072

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

Fires on
0.40% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 74%
  • chat 15%
  • agent 8%
  • worksheet 2%
  • python 1%

Makes the model say

n not have us rest share sweet make

Makes it not say

script Then trainedOnery flew<|think_start|>https

Strongest examples

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

  1. story 4.04
    for him to reach.⏎Tim asked his mom, "Can you help me get the cone
  2. story 4.02
    y the bird, for help. "Birdy, can you help me get this acorn
  3. story 3.95
    oven. You may not march like the ants, but you can help us make yummy food
  4. story 3.91
    't." The squirrel thought for a moment and said, "I can help you escape,
  5. story 3.89
    the tree, but the ice is too slippery. Can you help me?" The bird
  6. story 3.86
    , Tom. He was a strong man. "Tom, can you help me? There might
  7. story 3.83
    "Hi!" said the kitten. "I'm lost. Can you help me find my mom
  8. story 3.8
    string with the big knot. The squirrel said, "I can help you!" The squirrel
  9. agent 3.77
    running in OpenCode, a coding agent harness.<|system_end|><|user_start|>How can you help me?<|user_end|><|assistant_start|>
  10. agent 3.77
    pass now.<|assistant_end|><|user_start|>How can you help me?<|user_end|><|assistant_start|>
  11. story 3.75
    oyster said, "Hi, boat! I am lost. Can you help me find my family
  12. chat 3.73
    I want that cake, but I can't steal it. Can you help me ask who it