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

Feature #1139

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

Fires on
0.26% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 55%
  • story 24%
  • chat 12%
  • python 8%
  • worksheet 1%

Makes the model say

long but likeu andn ItOnce

Makes it not say

<|arg|>oldString<|think_start|> `])<|tool_call_start|>=**/

Strongest examples

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

  1. story 1.98
    happy and thanked Joy.⏎Joy learned a valuable lesson that
  2. agent 1.86
    model trained from scratch on a laptop. I can list, read, run, create, edit
  3. agent 1.85
    <|system_end|><|user_start|>Please write a script hello.py that prints "Hello, Sam!".<|user_end|><|assistant_start|><|tool_call_start|>
  4. agent 1.84
    that only knows small Python projects. I can list, read, run, create, edit
  5. story 1.84
    was testing their bravery and wanted to teach them a valuable lesson.
  6. agent 1.83
    "Write a script hi.py that prints "Hello, Lia!"."<|user_end|><|assistant_start|>
  7. agent 1.83
    model trained from scratch on a laptop. I can list, read, run, create, edit
  8. agent 1.82
    model trained from scratch on a laptop. I can list, read, run, create, edit
  9. agent 1.81
    .<|system_end|><|user_start|>Write a script greeting.py that prints "Hello, Sam!".<|user_end|><|assistant_start|><|tool_call_start|>
  10. agent 1.81
    model trained from scratch on a laptop. I can list, read, run, create, edit
  11. agent 1.8
    that only knows small Python projects. I can list, read, run, create, edit
  12. story 1.79
    their tea party, laughing and having fun, knowing they learned a valuable lesson.