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

Feature #2235

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

Fires on
0.10% of tokens
Mostly in
story
Layer
after block 4

Where it fires

  • story 44%
  • python 27%
  • agent 22%
  • chat 4%
  • worksheet 2%

Makes the model say

h len([ somee couldnMaybeul

Makes it not say

edage trainedutilsMITassert<|tool_call_end|> fact

Strongest examples

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

  1. python 1.73
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  2. python 1.71
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  3. python 1.7
    vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  4. agent 1.68
    ⏎def vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  5. python 1.66
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  6. python 1.63
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  7. python 1.62
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  8. python 1.62
    ⏎def vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  9. python 1.6
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  10. python 1.6
    ⏎def vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  11. python 1.59
    count_vowels(s):⏎ k = 0⏎ for c in s:⏎ if
  12. story 1.58
    day, his mom made him a new dish. It was called cauliflower.