Skip to content
mini-lab
Features of prelude-2

Feature #2351

Fires on “<|assistant_start|>” (not labeled yet: read the examples)

Fires on
2.69% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 90%
  • python 5%
  • chat 3%
  • story 2%
  • worksheet 0%

Makes the model say

RenamedAddedSetoldString definesCreated Python`,

Makes it not say

allter went first a She very He

Strongest examples

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

  1. agent 2.2
    return [n * 2 for n in nums]<|tool_end|><|assistant_start|>`smallest(nums
  2. story 2.13
    always remind her of the beautiful tree, and what she can achieve when she puts
  3. agent 2.1
    passed⏎<|tool_end|><|assistant_start|>`test_words
  4. agent 2.09
    return s.lower()<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  5. agent 2.09
    return s.lower()<|tool_end|><|assistant_start|>`text.py
  6. agent 2.05
    ⏎<|tool_end|><|assistant_start|>Fixed `size`
  7. agent 2.05
    ⏎<|tool_end|><|assistant_start|>Fixed `fact`
  8. agent 2.03
    contains(items, x):⏎<|tool_end|><|assistant_start|>`contains` is
  9. agent 2.03
    return s[-1]<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  10. agent 1.97
    <|assistant_end|><|tool_start|>Edited number_utils.py (1 replacement)<|tool_end|><|assistant_start|>Added `average(a
  11. agent 1.96
    ") == 5⏎<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>
  12. agent 1.93
    ⏎10: return x - 1<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>