Skip to content
mini-lab
Features of prelude-2

Feature #387

Fires on “2” (not labeled yet: read the examples)

Fires on
2.05% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 77%
  • chat 20%
  • agent 2%
  • story 1%
  • python 0%

Makes the model say

125067 owl3

Makes it not say

Python high around his}min sun each

Strongest examples

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

  1. chat 4.2
    9+6+1=16, 1691<|think_end|>The answer is
  2. worksheet 4.16
    8: 7+8=15, 5402⏎4+3
  3. worksheet 4.14
    18: 2+8=10, 082⏎06+
  4. worksheet 4.08
    9+8+1=18, 8162⏎8+0
  5. worksheet 4.06
    : 9+8+1=18, 864⏎61+
  6. worksheet 4.05
    0+9+1=10, 0129⏎0+5
  7. chat 4.04
    9+2+1=12, 1265<|think_end|>126
  8. worksheet 4.04
    9+9+1=19, 1963⏎So 9
  9. worksheet 4.02
    2+9+1=12, 2042⏎0+9
  10. chat 4.02
    8+8+1=17, 1706<|think_end|>The answer is
  11. worksheet 3.99
    : 9+9+1=19, 941⏎0+4
  12. worksheet 3.99
    5+8+1=14, 4704⏎4+0