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

Feature #11

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

Fires on
0.59% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 82%
  • story 7%
  • chat 5%
  • python 3%
  • worksheet 2%

Makes the model say

contentNomd<|tool_end|> his to dropped OpenCode

Makes it not say

many else)`item storiesave am In

Strongest examples

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

  1. worksheet 2.74
    9 + 252 = 261.⏎Hey, what's 4
  2. worksheet 2.69
    879 = 10564.⏎Hey, what's 2
  3. worksheet 2.53
    7 + 73 = 8880.⏎Hey, what's 4
  4. worksheet 2.42
    8 + 1 = 40379.⏎Hey, what's 4
  5. worksheet 2.4
    1537 = 1556.⏎Hey, what's 7
  6. worksheet 2.4
    986 + 4 = 990.⏎Hey, what's 2
  7. worksheet 2.37
    8547 = 9536.⏎Hey, what's 1
  8. worksheet 2.37
    9652 = 29657.⏎Hey, what's 6
  9. chat 2.36
    8 + 23 = 31.<|bos|><|user_start|>Hey, what's 2
  10. worksheet 2.36
    6 + 399 = 595.⏎Hey, what's 2
  11. chat 2.34
    95<|think_end|>The answer is 95.<|assistant_end|><|bos|><|user_start|>Hey, what's 8
  12. worksheet 2.34
    8124 = 68132.⏎Hey, what's 4