Skip to content
mini-lab
Features of prelude-2

Feature #1683

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

Fires on
0.38% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 76%
  • story 20%
  • chat 2%
  • python 1%
  • worksheet 1%

Makes the model say

files list you people<|assistant_end|><|tool_start|>s double

Makes it not say

Don upon# won8ustisher

Strongest examples

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

  1. story 2.49
    girl named Emma wanted to enter the big room where her mom was watching a film. She
  2. story 2.46
    One night, Jane decided to go to the stadium to watch the game. As
  3. agent 2.46
    most recent call last):⏎ File "main.py", line 4, in <module
  4. story 2.41
    Every day, he would sit in front of the TV and watch his favorite cartoon about a funny
  5. agent 2.37
    call last):⏎ File "test_arith.py", line 5, in <module
  6. agent 2.36
    call last):⏎ File "test_calc.py", line 3, in <module
  7. story 2.33
    was a gloomy room. The room had old furniture that no one liked. One
  8. agent 2.33
    call last):⏎ File "test_listlib.py", line 4, in <module
  9. story 2.32
    . Her mom said, "We can bake the cake after we watch a movie
  10. agent 2.31
    call last):⏎ File "test_utils.py", line 3, in <module
  11. agent 2.3
    ):⏎ File "test_textlib.py", line 6, in <module
  12. story 2.27
    that he had come to the beach to bury the violin so that no one else could keep