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

Feature #2899

Fires on “4” (not labeled yet: read the examples)

Fires on
0.33% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 52%
  • worksheet 27%
  • chat 13%
  • story 6%
  • python 2%

Makes the model say

<|user_end|><|tool_end|> aboutall [expression<|assistant_end|> sentence

Makes it not say

They!" two all?" tree they."

Strongest examples

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

  1. worksheet 3.58
    43.⏎67
  2. worksheet 3.58
    43, you get
  3. chat 3.51
    44+25:
  4. chat 3.51
    4+82:
  5. worksheet 3.51
    45⏎So
  6. agent 3.51
    4: print(mean
  7. worksheet 3.51
    47+03
  8. chat 3.51
    4.<|assistant_end|><|user_start|>what
  9. chat 3.51
    495<|assistant_end|><|bos|>
  10. chat 3.51
    4=11,
  11. chat 3.51
    4<|tool_end|><|assistant_start|>The answer
  12. agent 3.51
    4: acc =