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

Feature #324

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

Fires on
0.35% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 79%
  • python 14%
  • story 5%
  • chat 1%
  • worksheet 1%

Makes the model say

al late+def doneer0 False

Makes it not say

wereies]]`. held she they wished

Strongest examples

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

  1. python 1.77
    ⏎code!⏎ROBOT⏎```<|assistant_end|><|bos|>Bug: `average_
  2. story 1.56
    ball he had never seen before.⏎Tom
  3. agent 1.52
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  4. agent 1.52
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  5. agent 1.52
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  6. agent 1.52
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  7. agent 1.5
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  8. agent 1.5
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  9. agent 1.5
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  10. agent 1.5
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  11. agent 1.49
    bfetch, websearch, write, execute<|system_end|><|system_start|>
  12. agent 1.49
    bfetch, websearch, write, execute<|system_end|><|system_start|>