Skip to content
mini-lab
Features of prelude-2

Feature #4031

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

Fires on
1.11% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 97%
  • story 2%
  • chat 0%
  • worksheet 0%
  • python 0%

Makes the model say

inside around thereaten away`.[::

Makes it not say

t moral unexpectedvowels sum She Max li

Strongest examples

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

  1. agent 2.96
    , subagent, webfetch, websearch, write,
  2. agent 2.95
    , subagent, webfetch, websearch, write,
  3. agent 2.94
    , subagent, webfetch, websearch, write,
  4. agent 2.94
    , subagent, webfetch, websearch, write,
  5. agent 2.93
    , subagent, webfetch, websearch, write,
  6. agent 2.93
    , subagent, webfetch, websearch, write,
  7. agent 2.93
    , subagent, webfetch, websearch, write,
  8. agent 2.93
    , subagent, webfetch, websearch, write,
  9. agent 2.93
    , subagent, webfetch, websearch, write,
  10. agent 2.92
    , subagent, webfetch, websearch, write,
  11. agent 2.92
    , subagent, webfetch, websearch, write,
  12. agent 2.92
    , subagent, webfetch, websearch, write,