Skip to content
mini-lab
Features of prelude-2

Feature #3281

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

Fires on
0.82% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 87%
  • story 6%
  • python 4%
  • worksheet 1%
  • chat 1%

Makes the model say

But<|tool_call_end|>demo function trainediceay fact

Makes it not say

Tim smallest big gets took join bush add

Strongest examples

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

  1. story 3.07
    went to the park to do some exercise. He loved
  2. agent 3.04
    fetch, websearch, write, execute<|system_end|><|system_start|>You
  3. agent 3.04
    fetch, websearch, write, execute<|system_end|><|system_start|>You
  4. agent 3.04
    fetch, websearch, write, execute<|system_end|><|system_start|>You
  5. agent 3.04
    fetch, websearch, write, execute, gith
  6. agent 3.04
    fetch, websearch, write, execute, conte
  7. agent 2.99
    websearch, write, execute<|system_end|><|system_start|>You
  8. agent 2.99
    websearch, write, execute<|system_end|><|system_start|>You
  9. agent 2.99
    websearch, write, execute<|system_end|><|system_start|>You
  10. agent 2.99
    websearch, write, execute<|system_end|><|system_start|>You
  11. agent 2.99
    websearch, write, execute<|system_end|><|system_start|>You
  12. agent 2.98
    fetch, websearch, write, execute<|system_end|><|system_start|>You