Skip to content
mini-lab
Features of prelude-2

Feature #3109

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

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

Where it fires

  • agent 66%
  • story 24%
  • python 4%
  • chat 4%
  • worksheet 2%

Makes the model say

new):red)` fun instead` double

Makes it not say

we([listingb seekpython We

Strongest examples

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

  1. agent 2.55
    <|assistant_end|><|tool_start|>Read file mathlib.py, lines 1-9⏎1:
  2. agent 2.52
    <|assistant_end|><|tool_start|>Read file number_utils.py, lines 1-8⏎1:
  3. agent 2.49
    <|assistant_end|><|tool_start|>Read file words.py, lines 1-6⏎1:
  4. agent 2.48
    <|assistant_end|><|tool_start|>Read file greetings.py, lines 1-14⏎1
  5. agent 2.46
    AI agent running in OpenCode, a coding
  6. agent 2.46
    AI agent running in OpenCode, a coding
  7. agent 2.46
    AI agent running in OpenCode, a coding
  8. agent 2.46
    AI agent running in OpenCode, a coding
  9. agent 2.38
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Read file run.py, lines 1-14⏎1
  10. agent 2.36
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Read file hello.py, lines 1-4⏎1:
  11. agent 2.35
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Read file arith.py, lines 1-5⏎1:
  12. agent 2.35
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Read file greetings.py, lines 1-10⏎1