Skip to content
mini-lab
Features of prelude-2

Feature #2792

Fires on “1” (not labeled yet: read the examples)

Fires on
2.71% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 42%
  • agent 30%
  • chat 19%
  • python 7%
  • story 2%

Makes the model say

10Helocalculatoraei2Runvowels

Makes it not say

, r missed s can / at sad

Strongest examples

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

  1. python 3.5
    1))⏎⏎$ python3 main.py⏎tigertiger⏎11⏎$ python
  2. python 3.07
    - 1⏎⏎$ python3 app.py⏎5⏎11⏎$ python
  3. worksheet 3.07
    toys and found 6815 more. Now Tim has 13926
  4. worksheet 3.06
    and found 71111 more. Now Ben has 15431
  5. worksheet 2.96
    1556⏎So 19 + 1
  6. python 2.91
    ⏎ return a - b⏎⏎$ python3 hello.py⏎11⏎-5
  7. worksheet 2.91
    and found 51615 more. Now Max has 10991
  8. python 2.89
    , 7, 7, 1, 0])⏎1⏎⏎`repeat
  9. chat 2.89
    How many books does Max have now?<|user_end|><|assistant_start|>The answer is 100.<|assistant_end|>
  10. worksheet 2.88
    and found 59685 more. Now Ben has 15264
  11. worksheet 2.88
    balls and found 8035 more. Now Sue has 17259
  12. worksheet 2.87
    So 17 + 1