Skip to content
mini-lab
Features of prelude-2

Feature #3827

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

Fires on
0.93% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 49%
  • python 31%
  • worksheet 8%
  • story 6%
  • chat 6%

Makes the model say

= longer anually heraynder both

Makes it not say

⏎ ning do Do (ave Try count

Strongest examples

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

  1. python 4.97
    in nums]⏎⏎⏎def add_all(nums):⏎ result = 0⏎
  2. python 4.82
    b⏎⏎⏎def factorial(n):⏎ result = 1⏎
  3. agent 4.51
    , " + name + "!"⏎⏎⏎def fact(n):⏎ result = 1⏎
  4. agent 4.47
    then run script.py<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=NUMBERS =<|tool_call_end|><|assistant_end|><|tool_start|>
  5. python 4.44
    == 0]⏎⏎⏎def add_all(nums):⏎ result = 0⏎
  6. agent 4.44
    NUMBERS = [9, 7, 2]<|arg|>newString=NUMBERS = [1,
  7. python 4.43
    demo.py⏎from listlib import total, min_item⏎⏎NUMBERS = [9,
  8. agent 4.43
    (items)⏎⏎⏎def sum_list(nums):⏎ result = 0⏎
  9. python 4.41
    ⏎# listlib.py⏎def add_all(nums):⏎ result = 0⏎
  10. agent 4.36
    and run the script please<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=VALUES =<|tool_call_end|><|assistant_end|><|tool_start|>
  11. python 4.34
    ⏎from listlib import evens, count⏎from ops import minus⏎⏎NUMBERS = [5,
  12. agent 4.33
    , 4]))⏎⏎⏎def add_all(nums):⏎ result = 0⏎