Skip to content
mini-lab
Features of prelude-2

Feature #1490

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

Fires on
0.91% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 55%
  • python 37%
  • story 4%
  • worksheet 2%
  • chat 2%

Makes the model say

surprise two b longer, smell familydef

Makes it not say

nderprint len happenang<|system_start|>⏎ climb

Strongest examples

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

  1. agent 4.02
    a⏎⏎⏎def average(a, b):⏎ return (a + b) /
  2. agent 3.69
    b⏎⏎⏎def mean(a, b):⏎ return (a + b) /
  3. python 3.65
    `return a + b / 2` with `return (a + b) /
  4. agent 3.63
    2⏎⏎⏎def average(a, b):⏎ return (a + b) /
  5. agent 3.61
    ()⏎⏎⏎def mean(a, b):⏎ return (a + b) /
  6. agent 3.6
    )⏎⏎⏎def mean(a, b):⏎ return (a + b) /
  7. agent 3.59
    0⏎⏎⏎def average(a, b):⏎ return (a + b) /
  8. python 3.55
    (a + b) / 3` with `return (a + b) /
  9. agent 3.52
    py (1 replacement)<|tool_end|><|assistant_start|>Added `min_of(a, b)` to `
  10. agent 3.52
    .py` defines 2 functions: `max_of(a, b)` and `
  11. agent 3.51
    py (1 replacement)<|tool_end|><|assistant_start|>Added `sum_two(a, b)` to `
  12. agent 3.51
    return a + b / 2` instead of `return (a + b) /