Skip to content
mini-lab
Features of prelude-2

Feature #122

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

Fires on
3.22% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 48%
  • chat 32%
  • python 13%
  • story 4%
  • worksheet 3%

Makes the model say

in to for ideaitems k can //

Makes it not say

Thank moral<|assistant_end|>R thankful<|bos|> proud thanked

Strongest examples

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

  1. agent 4.13
    in OpenCode, a coding agent harness.<|system_end|><|user_start|>"Please run the tests and fix
  2. agent 4.11
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Please app.py cras
  3. agent 3.87
    agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Please run hello.py for me
  4. agent 3.79
    running in OpenCode, a coding agent harness.<|system_end|><|user_start|>"Explain hello from words.py
  5. agent 3.76
    OpenCode, a coding agent harness.<|system_end|><|user_start|>"Run run.py."<|user_end|>
  6. agent 3.67
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Please run the tests and fix
  7. agent 3.65
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Run run.py and fix
  8. agent 3.57
    AI agent running in OpenCode, a coding agent harness.<|system_end|><|user_start|>Please main.py cras
  9. agent 3.55
    and `final_letter(s)`.<|assistant_end|><|user_start|>Execute run.py.<|user_end|>
  10. agent 3.55
    <|assistant_end|><|user_start|>What's the output of run.py?<|user_end|>
  11. agent 3.51
    )` and `how_many(items)`.<|assistant_end|><|user_start|>Can you run hello.py?<|user_end|>
  12. agent 3.5
    py` with `mean_list(nums)`.<|assistant_end|><|user_start|>Can you run script.py?