Skip to content
mini-lab
Features of prelude-2

Feature #2520

Fires on “3” (not labeled yet: read the examples)

Fires on
0.07% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 65%
  • story 29%
  • chat 4%
  • agent 1%
  • python 1%

Makes the model say

ome timessum off ==<|tool_call_end|>lemon waved

Makes it not say

[::- python[- ever⏎⏎⏎ alwayson

Strongest examples

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

  1. worksheet 1.6
    7⏎So 6 + 58681 = 58687
  2. worksheet 1.39
    add 7 and 37648, you get 37655
  3. worksheet 1.26
    and found 39868 more. Now Tim has 39876
  4. worksheet 1.24
    4⏎So 7 + 34627 = 34634
  5. worksheet 1.23
    zoo.<|assistant_end|><|bos|>7 plus 44311 equals 44318
  6. worksheet 1.19
    and found 41886 more. Now Tom has 41891
  7. worksheet 1.08
    and found 68993 more. Now Anna has 69000
  8. worksheet 1.08
    ⏎96+47: 6+7+1=14, 4
  9. worksheet 1.07
    67.⏎7 plus 21509 is 21516
  10. story 1.07
    it's better to just relax and appreciate the things around him.<|bos|>
  11. worksheet 1.06
    `.<|assistant_end|><|bos|>57 plus 59600 is 59657
  12. worksheet 1.05
    and found 94109 more. Now Tom has 94117