Skip to content
mini-lab
Features of prelude-2

Feature #1176

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

Fires on
0.29% of tokens
Mostly in
python
Layer
after block 4

Where it fires

  • python 48%
  • worksheet 28%
  • chat 13%
  • story 8%
  • agent 2%

Makes the model say

(Added Theirins thingcontentone pile

Makes it not say

bettered saidation0 blue "D

Strongest examples

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

  1. worksheet 3.61
    26 is 104550.<|bos|>The sum of 83 and
  2. worksheet 3.56
    a great adventure because of the special medicine.<|bos|>The sum of 5 and
  3. worksheet 3.47
    0 + 46 = 46.<|bos|>The sum of 6 and
  4. python 3.28
    ⏎⏎`count_items` returns the number of items. For example, `count_
  5. worksheet 3.26
    8 more. Now Anna has 98 cookies.<|bos|>The sum of 76 and
  6. worksheet 3.23
    5 and 1, you get 36.⏎The sum of 300
  7. worksheet 3.23
    answer is 7.<|assistant_end|><|bos|>The sum of 277
  8. worksheet 3.22
    5<|think_end|>The answer is 695.<|assistant_end|><|bos|>The sum of 487
  9. python 3.21
    ⏎```⏎⏎`count` returns the number of items. For example, `count([
  10. worksheet 3.2
    44, you get 1234.⏎The sum of 619
  11. worksheet 3.18
    899 is 75928.<|bos|>The sum of 883
  12. python 3.17
    ⏎⏎`how_many` returns the number of items. For example, `how_