Skip to content
mini-lab
Features of prelude-2

Feature #21

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

Fires on
0.42% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 67%
  • python 10%
  • chat 10%
  • story 8%
  • worksheet 5%

Makes the model say

happy learnedre loved head love lastE

Makes it not say

=} % So⏎ uponexpressioncommand

Strongest examples

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

  1. python 3.04
    sum_list` should add up all the numbers in nums
  2. python 2.96
    ⏎ return result⏎```⏎⏎`sum_list` adds up all the numbers in nums
  3. agent 2.88
    all tests passed⏎<|tool_end|><|assistant_start|>`test_words.py` passes: all tests passed.
  4. agent 2.87
    all tests passed⏎<|tool_end|><|assistant_start|>`test_greetings.py` passes: all tests passed.
  5. agent 2.86
    all tests passed⏎<|tool_end|><|assistant_start|>`test_utils.py` passes: all tests passed.
  6. agent 2.84
    all tests passed⏎<|tool_end|><|assistant_start|>`test_words.py` passes: all tests passed.
  7. agent 2.8
    all tests passed⏎<|tool_end|><|assistant_start|>`test_words.py` passes: all tests passed.
  8. agent 2.79
    all tests passed⏎<|tool_end|><|assistant_start|>`test_items.py` passes: all tests passed.
  9. agent 2.79
    all tests passed⏎<|tool_end|><|assistant_start|>`test_lists.py` passes: all tests passed.
  10. agent 2.79
    all tests passed⏎<|tool_end|><|assistant_start|>`test_words.py` passes: all tests passed.
  11. agent 2.78
    all tests passed⏎<|tool_end|><|assistant_start|>`test_utils.py` passes: all tests passed.
  12. agent 2.77
    all tests passed⏎<|tool_end|><|assistant_start|>`test_tools.py` passes: all tests passed.