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

Feature #1741

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

Fires on
7.77% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 94%
  • story 2%
  • worksheet 2%
  • python 1%
  • chat 1%

Makes the model say

<|tool_call_start|>path<|arg|> sometimes not<|system_end|> defines't

Makes it not say

Once#One but")` OnehingBug

Strongest examples

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

  1. agent 3.71
    ]⏎7: ⏎8: ⏎9: NUMBERS = [1, 2
  2. agent 3.67
    ⏎7: ⏎8: ⏎9: TEXT = "cat"⏎1
  3. agent 3.66
    1⏎7: ⏎8: ⏎9: VALUES = [1, 8
  4. agent 3.56
    )⏎7: ⏎8: ⏎9: VALUES = [7, 0
  5. agent 3.56
    1: ⏎12: ⏎13: VALUES = [1, 6
  6. agent 3.56
    )⏎7: ⏎8: ⏎9: SCORES = [7, 7
  7. agent 3.54
    ⏎5: def exclaim(s):⏎6: return s + "!"<|tool_end|><|assistant_start|>`
  8. agent 3.53
    : ⏎4: ⏎5: def repeat(s, n):⏎6:
  9. agent 3.53
    stats import how_many⏎3: ⏎4: SCORES = [7, 0
  10. agent 3.52
    ⏎5: def factorial(n):⏎6: result = 1⏎7
  11. agent 3.51
    ⏎7: ⏎8: ⏎9: print(evens([5, 8
  12. agent 3.51
    ⏎7: ⏎8: ⏎9: print(doubled([4, 4