Skip to content
mini-lab
Log in Sign up
Features of prelude-2

Feature #1374

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

Fires on
0.72% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 39%
  • agent 38%
  • worksheet 13%
  • python 8%
  • story 2%

Makes the model say

so your,sny greetings0 long

Makes it not say

<|assistant_end|>patholdString<|tool_start|>ize way <|tool_end|>

Strongest examples

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

  1. worksheet 2.21
    <|user_end|><|assistant_start|>Happy to help!<|assistant_end|><|bos|>Mia had 0 coins and found 8 more
  2. worksheet 2.11
    're welcome!<|assistant_end|><|bos|>Sam had 24354 shells and found 19
  3. worksheet 1.94
    0 = 46.⏎Lucy had 54 shells and found 42
  4. worksheet 1.91
    99 = 103.⏎Ben had 5 shells and found 8 more
  5. worksheet 1.88
    = 1377⏎Ben had 41 apples and found 38
  6. worksheet 1.88
    Create hello.py script<|assistant_end|><|bos|>Tim had 357 coins and found 70
  7. worksheet 1.87
    166.<|assistant_end|><|bos|>Ben had 3512 toys and found 65
  8. worksheet 1.85
    The answer is 77.<|assistant_end|><|bos|>Tim had 0 cookies and found 9 more
  9. python 1.81
    b⏎```⏎⏎Bug: `includes` should check whether items contains x, but it
  10. worksheet 1.81
    4 = 1101.⏎Ben had 2 shells and found 23
  11. chat 1.81
    use the calculator.<|system_end|><|user_start|>Mia has 8 stickers. Mia gets 2 more stickers.
  12. worksheet 1.79
    0 = 4503.⏎Tom had 3 apples and found 1 more