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

Feature #1751

Fires on “1” (not labeled yet: read the examples)

Fires on
0.26% of tokens
Mostly in
python
Layer
after block 4

Where it fires

  • python 38%
  • story 27%
  • agent 19%
  • worksheet 12%
  • chat 3%

Makes the model say

],, gre Ben mess same<|tool_call_start|> living

Makes it not say

sooniceyinslies keptprintThey

Strongest examples

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

  1. python 6.58
    `has_item([4, 5, 5, 4], 4)`
  2. python 6.56
    assert has_item([1, 0, 2, 1], 9)
  3. python 6.19
    >>> has_item([1, 4, 5, 1], 4)
  4. python 6.03
    (has_item([6, 9, 5, 8], 5))
  5. python 6.03
    ⏎⏎>>> includes([2, 6, 3, 6, 5],
  6. python 5.94
    ⏎print(contains([5, 8, 7, 1], 7))
  7. python 5.85
    ⏎⏎>>> contains([2, 8, 7, 7], 1)
  8. python 5.8
    ⏎⏎assert contains([9, 0, 4, 2, 0],
  9. python 5.77
    assert has_item([5, 0, 6, 8, 2],
  10. python 5.73
    ⏎⏎>>> includes([0, 8, 6, 8, 8],
  11. python 5.66
    ⏎print(contains([9, 2, 5, 2], 2))
  12. python 5.63
    example, `contains([4, 4, 5, 0, 0],