Skip to content
mini-lab
Features of prelude-2

Feature #958

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

Fires on
1.22% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 98%
  • story 2%
  • worksheet 0%
  • chat 0%
  • python 0%

Makes the model say

<|assistant_end|>doubledate<|arg|>ing=**/*. likesSorry

Makes it not say

(_ + Sue ` she tests "

Strongest examples

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

  1. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  2. agent 4.03
    etch, websearch, write, execute, context
  3. agent 4.03
    etch, websearch, write, execute, lsp<|system_end|>
  4. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  5. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  6. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  7. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  8. agent 4.03
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  9. agent 4.01
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  10. agent 4.01
    etch, websearch, write, execute<|system_end|><|system_start|>You are
  11. agent 4.01
    etch, websearch, write, execute, context
  12. agent 4.0
    etch, websearch, write, execute<|system_end|><|system_start|>You are