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

Feature #3853

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

Fires on
0.18% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 77%
  • story 17%
  • chat 5%
  • python 1%
  • worksheet 0%

Makes the model say

( worked name never not knocked return list

Makes it not say

bighe him your offaypl into

Strongest examples

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

  1. story 3.06
    And when the sun went down, they all went home with happy memories of the shining
  2. agent 2.98
    websearch, write, execute, memory<|system_end|><|system_start|>You
  3. story 2.96
    went home, carrying her lovely jacket, with happy memories of a fun
  4. agent 2.89
    websearch, write, execute, memory, tod
  5. agent 2.87
    websearch, write, execute, memory, conte
  6. agent 2.83
    websearch, write, execute, memory, gith
  7. agent 2.79
    websearch, write, execute, memory, tod
  8. agent 2.78
    execute, github_search, memory<|system_end|><|system_start|>You
  9. agent 2.78
    websearch, write, execute, memory, brow
  10. agent 2.78
    websearch, write, execute, memory<|system_end|><|system_start|>You
  11. agent 2.78
    execute, context7_query, memory<|system_end|><|system_start|>You
  12. agent 2.77
    execute, context7_query, memory<|system_end|><|system_start|>You