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

Feature #1109

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

Fires on
2.57% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 93%
  • python 3%
  • story 2%
  • chat 1%
  • worksheet 1%

Makes the model say

⏎ themreplaceAlldanguckher hop

Makes it not say

wereAdded feel table= toydef import

Strongest examples

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

  1. agent 4.81
    ⏎⏎⏎number_utils.py:⏎ Line 7: def raise_to(a
  2. agent 4.68
    r"⏎⏎⏎textlib.py:⏎ Line 1: def final_letter(s
  3. agent 4.64
    VALUES))⏎⏎⏎items.py:⏎ Line 1: def add_all(nums
  4. agent 4.57
    ⏎⏎⏎math_utils.py:⏎ Line 7: def is_positive(x
  5. agent 4.55
    anana"⏎⏎⏎words.py:⏎ Line 5: def repeat(s, n
  6. agent 4.54
    ⏎⏎⏎math_utils.py:⏎ Line 7: def is_even(x
  7. agent 4.52
    file calc.py, lines 1-2⏎1: def next_number(x
  8. agent 4.48
    3 matches⏎lib.py:⏎ Line 1: def mean(a, b
  9. agent 4.41
    file listlib.py, lines 1-2⏎1: def mean_list(nums
  10. agent 4.4
    3 matches⏎calc.py:⏎ Line 9: def mean(a, b
  11. agent 4.38
    ⏎number_utils.py:⏎ Line 1: def smaller(a, b
  12. agent 4.36
    test_ops.py:⏎ Line 1: from ops import double, larger, negate