Skip to content
mini-lab
Features of prelude-2

Feature #1092

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

Fires on
0.67% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 48%
  • agent 23%
  • chat 14%
  • python 10%
  • story 4%

Makes the model say

for was more,9 looked not listened

Makes it not say

⏎ examplecountassert<|bos|>sumOfmany

Strongest examples

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

  1. python 2.47
    python⏎def only_even(nums):⏎ return [n for n in nums if
  2. python 2.46
    ```python⏎def only_even(nums):⏎ return [n for n in nums if
  3. python 2.42
    (nums):⏎ return [n for n in nums if
  4. python 2.4
    ```python⏎def keep_even(nums):⏎ return [n for n in nums if
  5. python 2.39
    ):⏎ return [n for n in nums if
  6. python 2.32
    ```python⏎def keep_even(nums):⏎ return [n for n in nums if
  7. agent 2.31
    n % 2 == 1]` instead of `return [n for n in nums if
  8. python 2.29
    result⏎⏎⏎def only_even(nums):⏎ return [n for n in nums if
  9. agent 2.29
    n % 2 == 1]` instead of `return [n for n in nums if
  10. agent 2.27
    n % 2 == 1]` instead of `return [n for n in nums if
  11. python 2.27
    ```python⏎def only_even(nums):⏎ return [n for n in nums if
  12. agent 2.26
    n % 2 == 1]` instead of `return [n for n in nums if