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

Feature #3384

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

Fires on
0.56% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 72%
  • python 20%
  • story 3%
  • worksheet 3%
  • chat 2%

Makes the model say

Maxjsonou [dwif seek

Makes it not say

]) made small project tiny]` plus x

Strongest examples

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

  1. python 3.24
    from words import loud⏎⏎TEXT = "robot"⏎print
  2. agent 3.11
    <|arg|>oldString=WORD = "tiger"<|arg|>newString
  3. agent 3.0
    py<|arg|>oldString=TEXT = "sun"<|arg|>newString=TEXT = "robot"<|tool_call_end|><|assistant_end|>
  4. agent 2.99
    py<|arg|>oldString=TEXT = "cat"<|arg|>newString=TEXT = "apple"<|tool_call_end|><|assistant_end|>
  5. python 2.92
    ⏎ result = 0⏎ for n
  6. agent 2.85
    USER = "Eva"<|tool_call_end|>
  7. agent 2.84
    oldString=TXT = "music"<|arg|>newString=TEXT = "house"<|tool_call_end|><|assistant_end|>
  8. agent 2.81
    0, 3, 5]<|arg|>newString=SCORES = [4, 5
  9. agent 2.81
    7, 0, 7]<|arg|>newString=SCORES = [9, 9
  10. agent 2.79
    5, 4, 8]<|arg|>newString=SCORES = [2, 6
  11. agent 2.77
    VALUES = [9, 8
  12. agent 2.76
    py<|arg|>oldString=TEXT = "banana"<|arg|>newString=TEXT = "music"<|tool_call_end|><|assistant_end|>