Skip to content
mini-lab
Features of prelude-2

Feature #1652

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

Fires on
0.18% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 44%
  • worksheet 26%
  • story 12%
  • python 9%
  • chat 8%

Makes the model say

6 looked9 startedSorry8 strong wore

Makes it not say

== sentence filled]ont ballsasedise

Strongest examples

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

  1. story 1.49
    dy had a special job to do. He wanted to make an old automobile
  2. worksheet 1.29
    Python file: `main.py`.<|assistant_end|><|bos|>I need to know 7 + 3
  3. worksheet 1.27
    14, 8]⏎```<|assistant_end|><|bos|>I need to know 9790
  4. agent 1.26
    test_arith.py<|arg|>oldString=positive<|arg|>newString=is_positive
  5. agent 1.25
    <|arg|>path=script.py<|arg|>oldString=SIZE = 7<|arg|>newString=SIZE =
  6. agent 1.23
    path=hello.py<|arg|>oldString= SIZE = 4<|arg|>newString=SIZE =
  7. worksheet 1.22
    py` and `arith.py`.<|assistant_end|><|bos|>I need to know 65 +
  8. agent 1.21
    path=mathlib.py<|arg|>oldString= result = 0<|arg|>newString= result =
  9. worksheet 1.2
    acc + i⏎ return acc⏎```⏎<|bos|>I need to know 950 +
  10. worksheet 1.19
    ops.py`, line 1.<|assistant_end|><|bos|>I need to know 6 + 4
  11. agent 1.19
    =test_calc.py<|arg|>oldString=absolute<|arg|>newString=magnitude<|arg|>
  12. story 1.18
    ⏎Their mom was very sad and disappointed. She said they were naughty and rude