Skip to content
mini-lab
Features of prelude-2

Feature #635

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

Fires on
0.87% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 44%
  • story 38%
  • python 9%
  • chat 6%
  • worksheet 2%

Makes the model say

lenhsc!")yingscript<|tool_call_end|>v

Makes it not say

veryappy so Mom wished would say done

Strongest examples

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

  1. story 3.68
    Be careful, the grater is sharp." Tom nods and grates the cheese
  2. story 3.61
    my eggs." Lily and Tom were very surprised. They did not know gems were bird eggs
  3. story 3.6
    ⏎The rabbit hopped off at tremendous speed while the dog galloped after
  4. agent 3.54
    , websearch, write, execute, github_se
  5. agent 3.52
    , websearch, write, execute, github_se
  6. agent 3.52
    , websearch, write, execute, github_se
  7. story 3.51
    squirrel. The squirrel said, "Hello, princess! Will you gather nuts with me
  8. story 3.49
    himself soaring high in the sky.⏎The clumsy monkey shouted with glee and danced
  9. agent 3.49
    , websearch, write, execute, github_se
  10. story 3.49
    butterfly. She was so excited that she applauded with glee.⏎
  11. agent 3.44
    , websearch, write, execute, github_se
  12. agent 3.44
    , websearch, write, execute, github_se