Skip to content
mini-lab
Features of prelude-2

Feature #3523

Fires on “5” (not labeled yet: read the examples)

Fires on
1.22% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 50%
  • worksheet 42%
  • agent 4%
  • story 3%
  • python 1%

Makes the model say

"" returnpy: import as about

Makes it not say

$ge trained prelude Finallynewneog

Strongest examples

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

  1. chat 3.28
    05586.<|bos|><|user_start|>What's 5 + 500?<|user_end|>
  2. chat 3.22
    something sweet, with just one squeeze!<|bos|><|user_start|>2 + 7 = ?<|user_end|><|assistant_start|>
  3. chat 3.21
    <|bos|><|user_start|>1 + 8<|user_end|><|assistant_start|>The answer
  4. chat 3.15
    every answer with "Sure!".<|system_end|><|user_start|>What's 2 + 519?<|user_end|>
  5. chat 3.12
    a new friend for Sue.<|bos|><|user_start|>What's 2 + 5?<|user_end|><|assistant_start|><|think_start|>
  6. worksheet 3.11
    74 = 35269.⏎9 plus 5 is 14
  7. chat 3.06
    another story soon.<|assistant_end|><|user_start|>Hey, what's 6 plus 6?<|user_end|><|assistant_start|>The
  8. chat 3.06
    The answer is 96.<|assistant_end|><|bos|><|user_start|>Compute 4+6<|user_end|><|assistant_start|>The answer
  9. worksheet 3.05
    3 = 144022.⏎3 plus 9976 is
  10. chat 3.05
    ⏎And they both lived happily ever after.<|bos|><|user_start|>3 + 9<|user_end|><|assistant_start|><|think_start|>3
  11. chat 3.02
    waited and hadn't given up!<|bos|><|user_start|>6 + 4<|user_end|><|assistant_start|>The answer
  12. worksheet 3.02
    35 equals 13289.⏎6 plus 5260 is