Skip to content
mini-lab
Features of prelude-2

Feature #3990

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

Fires on
0.56% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 76%
  • story 11%
  • chat 9%
  • python 4%
  • worksheet 0%

Makes the model say

): carefulJ soon each firstdyar

Makes it not say

unexpected moral<|user_start|><|think_end|><|arg|> satmdenv

Strongest examples

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

  1. agent 2.02
    help<|user_end|><|assistant_start|>I can list, read, run, create, edit and fix the Python files
  2. agent 1.93
    "<|user_end|><|assistant_start|>I can list, read, run, create, edit and fix the Python files
  3. agent 1.91
    <|assistant_start|>Hey! I can list, read, run, create, edit and fix the Python files
  4. chat 1.9
    can tell you short stories for children and add numbers, step by step or with a calculator.
  5. chat 1.86
    . Try asking me for a story!<|assistant_end|><|user_start|>how tall is the Eiffel Tower
  6. agent 1.86
    <|assistant_start|>Hey! I can list, read, run, create, edit and fix the Python files
  7. agent 1.86
    a laptop. I can list, read, run, create, edit and fix the Python files
  8. agent 1.86
    a laptop. I can list, read, run, create, edit and fix the Python files
  9. agent 1.85
    a laptop. I can list, read, run, create, edit and fix the Python files
  10. chat 1.85
    can tell you short stories for children and add numbers, step by step or with a calculator.
  11. agent 1.85
    running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Read number_utils.py.<|user_end|>
  12. agent 1.83
    help<|user_end|><|assistant_start|>I can list, read, run, create, edit and fix the Python files