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

Feature #2652

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

Fires on
0.46% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 51%
  • agent 23%
  • story 10%
  • python 9%
  • worksheet 7%

Makes the model say

. sun were them re= win their

Makes it not say

defined<|think_end|> here$ >>>ge

Strongest examples

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

  1. agent 1.66
    cute<|system_end|><|system_start|>You
  2. agent 1.66
    cute<|system_end|><|system_start|>You
  3. story 1.66
    cing and rolling
  4. story 1.66
    cense.⏎The
  5. agent 1.66
    cute<|system_end|><|system_start|>You
  6. agent 1.66
    cute<|system_end|><|system_start|>You
  7. agent 1.66
    cute<|system_end|><|system_start|>You
  8. chat 1.52
    much is 5 + 1?<|user_end|><|assistant_start|>The answer is 6.<|assistant_end|><|user_start|>Add
  9. story 1.5
    . They have many toys and books in their drawers. Sometimes they share, sometimes they fight
  10. chat 1.5
    not sad anymore.<|assistant_end|><|bos|>
  11. chat 1.5
    not sad anymore.⏎But
  12. story 1.48
    room was not messy anymore.<|bos|>7