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

Feature #2427

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

Fires on
0.23% of tokens
Mostly in
chat
Layer
after block 4

Where it fires

  • chat 42%
  • agent 22%
  • python 18%
  • story 12%
  • worksheet 6%

Makes the model say

" story Python eyesned rolled<|tool_call_end|> whisper

Makes it not say

they rest'tchars he kind we too

Strongest examples

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

  1. agent 2.06
    ⏎def fact(n):⏎ result = 1⏎ for i in range(1,
  2. agent 2.02
    sum_to(n):⏎ acc = 0⏎ for i in range(n +
  3. python 2.01
    up_to(n):⏎ acc = 0⏎ for i in range(n +
  4. chat 1.96
    ?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  5. chat 1.95
    ?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  6. chat 1.94
    .<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  7. chat 1.93
    ?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  8. chat 1.92
    pasta<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  9. python 1.92
    up_to(n):⏎ acc = 0⏎ for i in range(n +
  10. chat 1.92
    .<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  11. chat 1.91
    ?<|user_end|><|assistant_start|>Sorry, I'm a tiny model: I can only tell short stories and add
  12. python 1.9
    up_to(n):⏎ acc = 0⏎ for i in range(n +