Skip to content
mini-lab
Features of prelude-2

Feature #2728

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

Fires on
0.30% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 39%
  • agent 31%
  • chat 20%
  • python 8%
  • story 2%

Makes the model say

Fixboarditemcommand= typo:`nums

Makes it not say

thisOnce they back new returningsum

Strongest examples

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

  1. chat 2.8
    Hi!<|user_end|><|assistant_start|>Hi there! Would you like a story,
  2. worksheet 2.67
    66⏎00+36: 0+6+1=7,
  3. worksheet 2.6
    0+1: 0+1+1=2,
  4. worksheet 2.57
    076⏎0+6: 0+6+1=7,
  5. worksheet 2.56
    58⏎00+66: 0+6+1=7,
  6. chat 2.53
    5.<|assistant_end|><|user_start|>hey there<|user_end|><|assistant_start|>Hi there! Would you like a story,
  7. worksheet 2.53
    6: 0+6+1=7,
  8. worksheet 2.5
    : 0+6+1=7,
  9. chat 2.49
    , 63⏎0+6: 0+6+1=7,
  10. chat 2.49
    !".<|system_end|><|user_start|>Hello<|user_end|><|assistant_start|>Sure! Hi there! Would you like a story,
  11. worksheet 2.48
    0, 0⏎6+0: 6+0+1=7,
  12. chat 2.44
    , 67⏎6+0: 6+0+1=7,