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

Feature #603

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

Fires on
0.22% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 41%
  • chat 34%
  • story 20%
  • agent 3%
  • python 2%

Makes the model say

HiThank thanked len hellodeflesHello

Makes it not say

now. both he stuck she, their

Strongest examples

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

  1. worksheet 2.89
    stars and found 84674 more. Now Max has 8468
  2. story 2.88
    , 607<|think_end|>607<|assistant_end|><|bos|>Ben and Lily liked to fish with their
  3. worksheet 2.83
    cookies and found 80754 more. Now Anna has 8083
  4. story 2.77
    `last_letter(s)`.<|assistant_end|><|bos|>Tim and Liz liked to play in the
  5. story 2.76
    and went back to the cave full and happy.<|bos|>Ben and Mia like to play with their
  6. worksheet 2.71
    py`.<|assistant_end|><|bos|>57 plus 59600 is 5965
  7. worksheet 2.7
    apples and found 66647 more. Now Lily has 6671
  8. story 2.67
    ` with `sum_to(n)`.<|assistant_end|><|bos|>Ben and Mia like to play with their
  9. worksheet 2.65
    books and found 94109 more. Now Tom has 9411
  10. worksheet 2.64
    stickers and found 68993 more. Now Anna has 6900
  11. worksheet 2.63
    The sum of 89 and 67985 is 6807
  12. story 2.6
    test_items.py⏎all tests passed⏎<|bos|>Tom and Lily like to collect things.