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

Feature #2651

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

Fires on
0.37% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 72%
  • chat 14%
  • story 9%
  • python 2%
  • worksheet 2%

Makes the model say

Yes beautiful python special matternamewords_

Makes it not say

⏎ were ==7 are4 pleasele

Strongest examples

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

  1. chat 3.41
    with the number only.<|system_end|><|user_start|>Add
  2. chat 3.41
    with the number only.<|system_end|><|user_start|>7
  3. chat 3.16
    <|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Sue
  4. chat 3.16
    <|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Find
  5. chat 3.1
    ⏎<|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Add
  6. chat 3.06
    198<|think_end|>198<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>what
  7. chat 3.03
    583<|think_end|>583<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Can
  8. story 3.01
    was very excited and started to count right away. He counted one, two, three, four
  9. agent 3.01
    ob, grep, question, read, shell, skill, subagent, we
  10. chat 3.01
    , or some numbers to add?<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Please
  11. chat 3.0
    88<|think_end|>1088<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Calculate
  12. chat 2.99
    ⏎```<|assistant_end|><|bos|><|system_start|>Answer with the number only.<|system_end|><|user_start|>Calculate