Skip to content
mini-lab
Features of prelude-2

Feature #1978

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

Fires on
0.89% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 36%
  • python 29%
  • worksheet 17%
  • chat 13%
  • story 5%

Makes the model say

_( exclamationoldString gets()` typoure

Makes it not say

c little a mom wind other⏎ m

Strongest examples

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

  1. python 3.41
    import keep_even⏎⏎print(keep_even
  2. worksheet 3.32
    rag she had been playing with was never seen again.<|bos|>Ben has 23 flowers
  3. worksheet 3.19
    + 12 = 80126.⏎Ben has 326
  4. python 3.13
    from text import count_vowels⏎from helpers import subtract⏎⏎print(count_vowels
  5. worksheet 3.1
    <|assistant_end|><|bos|>Tom has 872
  6. worksheet 3.09
    add 2 and 1, you get 3.<|bos|>Anna has 210
  7. python 3.05
    43<|think_end|>The answer is 43.<|assistant_end|><|bos|># opencode.json⏎{
  8. python 3.05
    I only know how to tell stories and add numbers.<|assistant_end|><|bos|># opencode.json⏎{
  9. worksheet 3.04
    7852 = 15316⏎Ben has 14 apples
  10. worksheet 3.04
    6 + 2 = 40858.⏎Tom has 743
  11. python 3.04
    26<|think_end|>The answer is 26.<|assistant_end|><|bos|># opencode.json⏎{
  12. worksheet 3.04
    154<|assistant_end|><|bos|>Tom has 243