Skip to content
mini-lab
Features of prelude-2

Feature #955

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

Fires on
0.63% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 83%
  • python 9%
  • chat 4%
  • story 3%
  • worksheet 1%

Makes the model say

about)) how short sea<|think_start|> homeall

Makes it not say

`" both From READMERunpython %Fix

Strongest examples

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

  1. python 3.17
    == True⏎assert contains([2, 3, 4], 3) == True
  2. python 3.14
    _item([1, 4, 5, 1], 4)⏎True
  3. python 3.09
    (contains([9, 2, 5, 2], 2))⏎⏎
  4. python 3.08
    (contains([1, 4, 1, 3], 3))⏎print
  5. python 3.06
    9, 0, 4, 2, 0], 2) == True
  6. python 3.06
    _item([1, 0, 2, 1], 9) == False
  7. python 3.03
    _item([6, 1, 9, 9], 2)⏎False
  8. python 3.03
    >>> contains([2, 8, 7, 7], 1)⏎False
  9. python 2.96
    For example, `contains([7, 0, 5], 5)` returns `
  10. python 2.95
    assert contains([7, 7, 7, 5], 0) == False
  11. python 2.94
    _item([4, 5, 5, 4], 4)` returns `
  12. python 2.93
    list⏎⏎assert includes([0, 8, 3], 8) == True