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

Feature #1234

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

Fires on
0.49% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 98%
  • story 2%
  • python 0%
  • chat 0%
  • worksheet 0%

Makes the model say

trained<|assistant_end|>rement checkonfig isoldString OpenCode

Makes it not say

have who " please= find need use

Strongest examples

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

  1. agent 2.74
    letter, exclaim⏎3: ⏎4: print(final_l
  2. agent 2.41
    3, 1, 4, 4]⏎4: print(head([5
  3. agent 2.41
    VALUES = [4, 0, 9]⏎4: print(how_many
  4. agent 2.41
    name + "?"⏎3: ⏎4: ⏎5: print(greet("Max
  5. agent 2.41
    8, 1, 1, 2]))⏎4: print(mean_ist
  6. agent 2.4
    0, 8, 4, 5]⏎4: print(only_even
  7. agent 2.39
    2: ⏎3: WORD = "moon"⏎4: print(first_letter
  8. agent 2.39
    NUMBERS = [4, 6, 6]⏎4: print(contains([2
  9. agent 2.39
    7, 5, 4, 8]⏎4: print(max_item
  10. agent 2.39
    8, 7, 2, 2]⏎4: print(first(VALUES
  11. agent 2.38
    NUMBERS = [2, 1, 9]⏎4: print(doubled(NUMBERS
  12. agent 2.38
    NUMBERS = [2, 9, 3]⏎4: print(tail(NUMBERS