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

Feature #74

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

Fires on
0.20% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 69%
  • story 17%
  • chat 8%
  • worksheet 4%
  • python 2%

Makes the model say

car hear(oodaring_ wered

Makes it not say

and each turns only⏎mainlytest

Strongest examples

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

  1. story 1.82
    mom. Mia saw an
  2. story 1.82
    mom. They saw many
  3. story 1.82
    mom, dad, and
  4. story 1.82
    mom is busy on the
  5. story 1.82
    mom and dad. They
  6. story 1.82
    mom came into the room
  7. story 1.82
    mom, and she came
  8. agent 1.79
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_calc.
  9. agent 1.78
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_items.
  10. agent 1.78
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_stats.
  11. agent 1.77
    <|tool_call_end|><|assistant_end|><|tool_start|>4⏎Traceback (most recent call last):⏎ File "main.py",
  12. agent 1.77
    .py<|tool_call_end|><|assistant_end|><|tool_start|>Traceback (most recent call last):⏎ File "test_items.