Skip to content
mini-lab
Features of prelude-2

Feature #983

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

Fires on
0.95% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 87%
  • worksheet 5%
  • story 4%
  • chat 3%
  • python 1%

Makes the model say

<|assistant_end|><|tool_call_start|>scommand wokeed our //

Makes it not say

very so but much2⏎ excited But

Strongest examples

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

  1. agent 2.41
    grep, question, read, shell, skill, subagent, webf
  2. agent 2.41
    grep, question, read, shell, skill, subagent, webf
  3. agent 2.4
    grep, question, read, shell, skill, subagent, webf
  4. agent 2.39
    grep, question, read, shell, skill, subagent, webf
  5. agent 2.39
    grep, question, read, shell, skill, subagent, webf
  6. agent 2.38
    grep, question, read, shell, skill, subagent, webf
  7. agent 2.38
    grep, question, read, shell, skill, subagent, webf
  8. agent 2.38
    grep, question, read, shell, skill, subagent, webf
  9. agent 2.37
    grep, question, read, shell, skill, subagent, webf
  10. agent 2.37
    grep, question, read, shell, skill, subagent, webf
  11. agent 2.37
    grep, question, read, shell, skill, subagent, webf
  12. agent 2.37
    grep, question, read, shell, skill, subagent, webf