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

Feature #1106

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

Fires on
0.50% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 70%
  • python 13%
  • chat 7%
  • worksheet 6%
  • story 4%

Makes the model say

then now soon watched later but`,replaceAll

Makes it not say

51047298

Strongest examples

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

  1. agent 2.85
    keep_even⏎⏎ Line 3: assert add_all([9, 6,
  2. agent 2.64
    assert count([8, 2,
  3. agent 2.52
    count_items, first⏎⏎ Line 5: assert first([8, 0,
  4. agent 2.51
    first, only_even⏎⏎ Line 3: assert first([5, 8,
  5. agent 2.49
    head, how_many⏎⏎ Line 3: assert head([6, 7,
  6. agent 2.49
    ⏎main.py:⏎ Line 3: VALUES = [1, 3,
  7. agent 2.46
    listlib import head, doubled⏎⏎ Line 3: assert head([2, 9,
  8. agent 2.43
    ⏎app.py:⏎ Line 3: VALUES = [7, 0,
  9. agent 2.41
    lsat([6, 3,
  10. agent 2.38
    all, mean_list⏎⏎ Line 3: assert count([7, 6,
  11. agent 2.37
    4, in <module>⏎ print(incldes([6, 2,
  12. agent 2.37
    count_items⏎⏎ Line 5: assert count_items([7, 8,