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

Feature #1459

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

Fires on
0.09% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 81%
  • chat 9%
  • story 8%
  • agent 1%
  • python 1%

Makes the model say

hadstageur only has⏎ smallest

Makes it not say

see prints until hiding Ben findreturn seen

Strongest examples

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

  1. worksheet 6.84
    14 + 96 = 6610.⏎Anna had 75
  2. worksheet 6.83
    58681 = 58687.⏎Sam has 16
  3. worksheet 6.72
    9 + 640 = 649.⏎Mia has 85
  4. worksheet 6.71
    19178 = 20700.⏎Lucy had 61
  5. worksheet 6.65
    2981 = 79424.⏎Mia has 8 books
  6. worksheet 6.64
    48698 = 49516.⏎Sam had 63
  7. worksheet 6.6
    4 + 69 = 36013.⏎Sam has 25
  8. worksheet 6.52
    465, you get 95100.⏎Mia has 82
  9. worksheet 6.5
    870, you get 15423.⏎Lucy had 84
  10. worksheet 6.39
    29 + 5 = 75834.⏎Mia had 35
  11. worksheet 6.38
    7249, you get 7256.⏎Lucy has 96
  12. worksheet 6.35
    + 4312 = 4396.⏎Sam had 7 stickers