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

Feature #487

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

Fires on
0.57% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 37%
  • python 25%
  • story 21%
  • chat 17%
  • worksheet 0%

Makes the model say

ofSamantcountpy same + feeling

Makes it not say

ightOnce]`.g walk ButFixedos

Strongest examples

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

  1. chat 2.36
    1 function: `tail(items)`.<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  2. chat 2.27
    you do them in the right order.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  3. chat 2.26
    pass now.<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  4. chat 2.25
    088<|think_end|>1088<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  5. chat 2.25
    , ready for another day of fun together.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  6. chat 2.24
    , they can turn out to be good.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  7. chat 2.22
    <|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  8. chat 2.22
    <|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  9. chat 2.21
    story, or some numbers to add?<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  10. chat 2.2
    22 is 1534.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  11. chat 2.2
    6 = 98926.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>
  12. chat 2.2
    she learned something new with her mom.<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>