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

Feature #2643

Fires on ā€œ nā€ (not labeled yet: read the examples)

Fires on
0.43% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 46%
  • python 45%
  • story 5%
  • worksheet 2%
  • chat 2%

Makes the model say

plus % times<|user_start|>): agtree >

Makes it not say

for onek waiting smell next waited seek

Strongest examples

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

  1. agent 6.89
    = return [n * 2 for n in nums]<|tool_call_end|>
  2. agent 6.7
    return [n * 2 for n in nums]`. The
  3. python 6.43
    in nums]` with `return [n * 2 for n in nums]`.āŽ
  4. agent 6.22
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  5. agent 6.05
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  6. agent 6.03
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  7. agent 6.02
    return [n * 2 for n in nums]<|tool_end|>
  8. agent 6.0
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  9. agent 5.97
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  10. agent 5.91
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  11. agent 5.89
    n in nums]` instead of `return [n * 2 for n in nums]`. The
  12. agent 5.88
    n in nums]` instead of `return [n * 2 for n in nums]`. The