Skip to content
mini-lab
Features of prelude-2

Feature #2332

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

Fires on
1.19% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 52%
  • python 30%
  • story 11%
  • worksheet 6%
  • chat 2%

Makes the model say

sopath a trained decided i time

Makes it not say

They They she example girl SheShe high

Strongest examples

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

  1. agent 2.47
    functions: `cubed(x)` and `max_of(a, b)`.<|assistant_end|><|bos|>
  2. agent 2.41
    result = result + n⏎13
  3. python 2.41
    result = result + n⏎ return result
  4. agent 2.29
    a coding agent harness.<|system_end|><|user_start|>Write a function only_even(nums) in core.
  5. story 2.25
    to hurt Anna. He likes her. He says, "Wait, Anna. Come
  6. story 2.24
    She was very excited.⏎Lucy's mom said, "Wait, we need to
  7. python 2.24
    safe during the flood.<|bos|>`only_even(nums)` keeps only
  8. python 2.23
    .py`, line 1.<|assistant_end|><|bos|>`keep_even(nums)` keeps only
  9. story 2.22
    out of the box and followed her. It said, "Wait, Lily. Don
  10. agent 2.22
    `add_one(x)` and `min_of(a, b)`.<|assistant_end|><|bos|>
  11. agent 2.2
    functions: `average_list(nums)` and `doubled(nums)`.<|assistant_end|><|user_start|>re
  12. agent 2.2
    ` with `max_of(a, b)`.<|assistant_end|><|bos|>