Skip to content
mini-lab
Features of prelude-2

Feature #1486

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

Fires on
0.40% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 83%
  • story 12%
  • python 3%
  • worksheet 2%
  • chat 1%

Makes the model say

from to by at playingAnnaate goodbye

Makes it not say

]`)) lots `()`")Do#

Strongest examples

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

  1. chat 1.59
    not sad anymore.<|assistant_end|><|bos|>
  2. chat 1.59
    not sad anymore.⏎But
  3. agent 1.55
    0]))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=print(f
  4. agent 1.51
    apple"))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=TEXT = "
  5. agent 1.51
    cat"))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=print(final
  6. agent 1.48
    SCORES))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString= print(how
  7. agent 1.47
    SCORES))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=SCORES = [
  8. agent 1.47
    6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=print(h
  9. agent 1.46
    (decrrement(6))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=run.py
  10. agent 1.46
    ar"))<|tool_end|><|assistant_start|><|tool_call_start|>edit<|arg|>path=main.py<|arg|>oldString=print(in
  11. chat 1.45
    6, 6⏎1
  12. python 1.45
    6, 1,