Skip to content
mini-lab
Features of prelude-2

Feature #3922

Fires on “<|assistant_end|>” (not labeled yet: read the examples)

Fires on
2.14% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 81%
  • chat 7%
  • worksheet 5%
  • python 4%
  • story 2%

Makes the model say

Hioo're is ItHello<|tool_call_end|>case

Makes it not say

instead both projectemember more Many the moral

Strongest examples

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

  1. agent 2.34
    hello.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def hello<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  2. agent 2.26
    do?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def whisper<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  3. agent 2.26
    smallest.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def smallest<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  4. agent 2.23
    hello."<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def hello<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  5. agent 2.22
    larger.<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def larger<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  6. agent 2.21
    whisper?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def whisper<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  7. agent 2.15
    tail?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def tail<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  8. agent 2.15
    return?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def whisper<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  9. agent 2.14
    return?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def halve<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  10. agent 2.13
    <|assistant_start|><|tool_call_start|>write<|arg|>path=tools.py<|arg|>content=def is_divisible(a
  11. agent 2.13
    exclaim?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def exclaim<|tool_call_end|><|assistant_end|><|tool_start|>Found 1
  12. agent 2.13
    do?<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def squared<|tool_call_end|><|assistant_end|><|tool_start|>Found 1