Skip to content
mini-lab
Features of prelude-2

Feature #869

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

Fires on
0.59% of tokens
Mostly in
agent
Layer
after block 4

Where it fires

  • agent 93%
  • chat 3%
  • python 2%
  • story 2%
  • worksheet 0%

Makes the model say

<|tool_call_end|> help Python]))!") forgivereplaceAllpath

Makes it not say

likedCan meThat⏎orAdd The

Strongest examples

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

  1. agent 3.27
    1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def min_item<|tool_call_end|><|assistant_end|>
  2. agent 3.15
    1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def backwards<|tool_call_end|><|assistant_end|><|tool_start|>Found
  3. agent 3.15
    is defined<|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def cube<|tool_call_end|><|assistant_end|><|tool_start|>Found
  4. agent 3.12
    code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def power<|tool_call_end|><|assistant_end|><|tool_start|>Found
  5. agent 3.05
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def minus<|tool_call_end|><|assistant_end|><|tool_start|>Found
  6. agent 2.98
    <|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def sum_to<|tool_call_end|><|assistant_end|>
  7. agent 2.98
    <|user_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def has_item<|tool_call_end|><|assistant_end|>
  8. agent 2.93
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def reverse<|tool_call_end|><|assistant_end|><|tool_start|>Found
  9. agent 2.92
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def magnitude<|tool_call_end|><|assistant_end|><|tool_start|>Found
  10. agent 2.91
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def decrement<|tool_call_end|><|assistant_end|><|tool_start|>Found
  11. agent 2.91
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def backwards<|tool_call_end|><|assistant_end|><|tool_start|>Found
  12. agent 2.91
    with code 1<|tool_end|><|assistant_start|><|tool_call_start|>grep<|arg|>pattern=def flip<|tool_call_end|><|assistant_end|><|tool_start|>Found