Skip to content
mini-lab
Features of prelude-2

Feature #3055

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

Fires on
0.41% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 65%
  • story 23%
  • python 8%
  • chat 3%
  • worksheet 0%

Makes the model say

() == thisher_ middle say This

Makes it not say

him,"`Bugcalculator sharpread long

Strongest examples

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

  1. chat 1.7
    : calculator<|system_end|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  2. story 1.69
    ince Tim did. She was not happy. She told Prince Tim, "
  3. chat 1.55
    .<|assistant_end|><|bos|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  4. agent 1.55
    it<|arg|>path=core.py<|arg|>oldString= return s.lower()<|arg|>newString=
  5. chat 1.55
    ?<|assistant_end|><|bos|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  6. story 1.51
    ?" she asked, taking the lid from them. "Why are you fighting over this?"
  7. chat 1.49
    .<|assistant_end|><|bos|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  8. chat 1.46
    patient.<|bos|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  9. agent 1.46
    7]) == 29⏎⏎⏎demo.py:⏎ Line 1: from
  10. chat 1.45
    0.<|assistant_end|><|user_start|>Tell me the latest sports scores.<|user_end|><|assistant_start|>
  11. agent 1.44
    , 0]) == 9⏎⏎⏎app.py:⏎ Line 1: from
  12. story 1.43
    ope. Hope loved to play with the furniture in her house