Skip to content
mini-lab
Features of prelude-2

Feature #2550

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

Fires on
0.34% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 76%
  • story 10%
  • python 7%
  • chat 5%
  • worksheet 2%

Makes the model say

demo=**/ also<|tool_call_end|>-angreplaceAll i

Makes it not say

``` butwords and PleaseOnce would around

Strongest examples

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

  1. agent 1.89
    `: it used `return s[1:]` instead of `return s[::-
  2. agent 1.79
    py`: it used `k = k + 2` instead of `k = k
  3. agent 1.75
    calc.py`: it used `return x * 3` instead of `return x *
  4. agent 1.74
    it used `return "Helo, " + name + "!"` instead of `return "Hello
  5. agent 1.73
    textlib.py`: it used `return s * 2` instead of `return s *
  6. agent 1.72
    `: it used `return len(s) - 1` instead of `return len(
  7. agent 1.69
    utils.py`: it used `return x - 2` instead of `return x -
  8. agent 1.69
    tools.py`: it used `return x + 1` instead of `return x -
  9. agent 1.69
    it used `return "Hello, " + name + "?"` instead of `return "Hello
  10. agent 1.68
    utils.py`: it used `return x * 3` instead of `return x *
  11. agent 1.67
    `return sum(nums) / 2` instead of `return sum(
  12. agent 1.67
    py`: it used `return a + b / 2` instead of `return (a