Skip to content
mini-lab
Features of prelude-2

Feature #1450

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

Fires on
1.07% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 53%
  • python 29%
  • chat 12%
  • story 3%
  • worksheet 3%

Makes the model say

Python mini laptoppytest:// "$Fixedalse

Makes it not say

1 are⏎ were<|bos|> count home am

Strongest examples

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

  1. python 4.14
    _list(SCORES))⏎⏎$ python3 app.py⏎6.0⏎<|bos|>
  2. python 3.74
    py`.<|assistant_end|><|bos|># tools.py⏎def
  3. python 3.63
    <|tool_end|><|assistant_start|>The answer is 1656.<|assistant_end|><|bos|># listlib.py⏎def
  4. python 3.61
    , 8, 3, 8, 2])⏎4.8⏎<|bos|>
  5. python 3.57
    home sad.<|bos|># utils.py⏎def
  6. python 3.56
    shadows on the wall. They were happy and safe.<|bos|># tools.py⏎def
  7. python 3.53
    add numbers.<|assistant_end|><|bos|># listlib.py⏎def
  8. python 3.52
    513<|think_end|>513<|assistant_end|><|bos|># tools.py⏎def
  9. python 3.5
    8, 4, 6, 8])` returns `6.2`.⏎
  10. python 3.49
    8, 6, 8]⏎22⏎<|bos|># utils.py⏎def
  11. python 3.49
    2670 = 92679.<|bos|># lib.py⏎def
  12. python 3.47
    list([0, 7, 5, 7])⏎4.75⏎