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

Feature #2856

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

Fires on
0.73% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 59%
  • story 13%
  • worksheet 12%
  • python 10%
  • chat 7%

Makes the model say

ferever<|tool_call_start|>us donnum<|assistant_start|>

Makes it not say

:hes sum am7⏎4ls

Strongest examples

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

  1. worksheet 3.65
    found 554 more. Now Anna has
  2. worksheet 3.64
    had 164 balls and found 7145 more. Now Tom has
  3. worksheet 3.62
    Sue has 99 stars. Sue gets 33 more stars. How many stars does
  4. worksheet 3.57
    had 5298 stars and found 664 more. Now Tim has
  5. worksheet 3.55
    3360 stars. Lily gets 555 more stars. How many stars does
  6. worksheet 3.54
    .<|bos|>Max had 2 stars and found 176 more. Now Max has
  7. chat 3.52
    has 87 stars. Anna gets 872 more stars. How many stars does
  8. agent 3.5
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  9. worksheet 3.46
    559 stickers. Sam gets 1731 more stickers. How many stickers does
  10. worksheet 3.46
    287 coins and found 76758 more. Now Max has
  11. agent 3.46
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  12. agent 3.45
    bfetch, websearch, write, execute<|system_end|><|system_start|>