Skip to content
mini-lab
Features of prelude-2

Feature #1344

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

Fires on
0.26% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 78%
  • story 15%
  • chat 5%
  • python 1%
  • worksheet 1%

Makes the model say

to " qu README Just up stwords

Makes it not say

} Pythonprint divisible Tom laterdouble

Strongest examples

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

  1. story 3.88
    time and never felt nosy again. Instead, he learned that it's better to just relax
  2. agent 3.53
    test_mathlib.py (3 replacements)<|tool_end|><|assistant_start|>Renamed `cubed` to `cubed_
  3. agent 3.49
    <|tool_start|>Edited lists.py (1 replacement)<|tool_end|><|assistant_start|>Renamed `head` to `head_
  4. story 3.48
    to celebrate their new friendship. Tom and Jerry learned that it is better to be friends and
  5. agent 3.44
    <|tool_start|>Edited mathlib.py (1 replacement)<|tool_end|><|assistant_start|>Renamed `odd` to `odd_
  6. chat 3.42
    and Max were not friends. They both learned that sometimes, it is better to share and be
  7. agent 3.41
    <|tool_start|>Edited lists.py (1 replacement)<|tool_end|><|assistant_start|>Renamed `count` to `how_
  8. agent 3.39
    <|tool_start|>Edited script.py (2 replacements)<|tool_end|><|assistant_start|>Renamed `factorial` to `my_
  9. story 3.38
    wise words, Pete and Tom learned the lesson that it is better for friends to share
  10. agent 3.36
    <|tool_start|>Edited listlib.py (1 replacement)<|tool_end|><|assistant_start|>Renamed `tail` to `my_
  11. story 3.35
    yarn. The moral of the story is that sometimes, it's better to share and be
  12. story 3.32
    clothes. In the end, they became friends and learned that it is better to work together than