Skip to content
mini-lab
Features of prelude-2

Feature #2907

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

Fires on
0.40% of tokens
Mostly in
python
Layer
after block 4

Where it fires

  • python 36%
  • chat 23%
  • agent 20%
  • story 16%
  • worksheet 5%

Makes the model say

defSoprint We us read So

Makes it not say

th<|user_start|> areThank2Good felt waiting

Strongest examples

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

  1. chat 3.12
    it in New York<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  2. chat 3.11
    I fix a flat tire?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  3. chat 3.1
    I fix a flat tire?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  4. chat 3.09
    <|user_start|>What does DNA stand for?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  5. chat 3.03
    is it in Berlin?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  6. chat 3.03
    the tallest mountain in the world?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  7. chat 3.01
    physics to me.<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  8. python 3.0
    ) == 1⏎assert sub_one(11) == 10⏎assert
  9. chat 2.99
    is the Eiffel Tower?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I
  10. chat 2.99
    I fix a flat tire?<|user_end|><|assistant_start|>Sorry, I'm a
  11. python 2.98
    max(nums)⏎⏎$ python3 demo.py⏎7⏎5⏎$ python
  12. chat 2.97
    is it in Sydney?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I