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

Feature #2045

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

Fires on
0.51% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 64%
  • python 25%
  • worksheet 7%
  • chat 2%
  • story 2%

Makes the model say

answer fundouble + trained` me you

Makes it not say

est songshes toldOneoreainname

Strongest examples

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

  1. agent 4.69
    4: acc = acc + 1⏎
  2. agent 4.49
    : acc = acc + i⏎5
  3. python 4.46
    sum of the numbers in nums, but it has `result = result + 1`.
  4. python 4.42
    numbers from 0 to n, but it has `acc = acc + 1`.
  5. python 4.29
    ):⏎ acc = acc + i⏎ return
  6. python 4.28
    ⏎ acc = acc + i⏎ return
  7. python 4.01
    1):⏎ acc = acc + i⏎ return
  8. python 3.99
    add up all the numbers in nums, but it has `result = result + 1`.
  9. agent 3.89
    ":⏎5: k = k + 1⏎
  10. agent 3.84
    result = result * i<|tool_call_end|><|assistant_end|>
  11. agent 3.74
    :⏎8: result = result + n⏎9
  12. python 3.73
    ⏎ for i in range(n + 1):⏎ acc = acc + i⏎ return