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

Feature #1377

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

Fires on
1.97% of tokens
Mostly in
worksheet
Layer
after block 4

Where it fires

  • worksheet 56%
  • chat 31%
  • python 6%
  • agent 4%
  • story 3%

Makes the model say

md⏎ ] Sue]))( " still

Makes it not say

milkGoodCreatedATheWhat<|tool_call_start|> fire

Strongest examples

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

  1. worksheet 4.14
    00 stickers.⏎The sum of 6 and 9 is 15.⏎
  2. chat 3.98
    + 93<|tool_call_end|><|assistant_end|>
  3. agent 3.98
    + 1<|tool_end|><|assistant_start|><|tool_call_start|>
  4. chat 3.98
    + 6<|tool_call_end|><|assistant_end|><|tool_start|>
  5. chat 3.98
    + 35<|tool_call_end|><|assistant_end|>
  6. chat 3.98
    + 9?<|user_end|><|assistant_start|>
  7. chat 3.98
    + 920 is
  8. chat 3.98
    + 619<|tool_call_end|>
  9. chat 3.98
    + 500?
  10. worksheet 3.98
    + 6248
  11. worksheet 3.98
    + 53 = ?
  12. worksheet 3.98
    + 73 =