Skip to content
mini-lab
Features of prelude-2

Feature #2558

Fires on “<|system_end|>” (not labeled yet: read the examples)

Fires on
0.48% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 83%
  • chat 16%
  • story 1%
  • worksheet 0%
  • python 0%

Makes the model say

replaceAll >ily glad continued sorryem cheered

Makes it not say

little ` big y but that p first

Strongest examples

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

  1. agent 3.49
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  2. agent 3.46
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  3. agent 3.45
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  4. agent 3.44
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  5. agent 3.43
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  6. agent 3.43
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  7. agent 3.42
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  8. agent 3.42
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  9. agent 3.42
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  10. agent 3.41
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  11. agent 3.4
    ch, websearch, write, execute<|system_end|><|system_start|>You are an
  12. agent 3.4
    ch, websearch, write, execute<|system_end|><|system_start|>You are an