Skip to content
The Chat Archive
Human-written demonstration (OpenAssistant volunteers)· OpenAssistant (human-written) OpenAssistant Conversations (oasst2)

Why does the string ' petertodd' cause GPT-derivative models to behave erratically?

In OpenAssistant data, “assistant” turns were written by human volunteers as demonstrations. They are not model outputs.
  1. User

    Why does the string ' petertodd' cause GPT-derivative models to behave erratically?

  2. Assistant

    The string "petertodd" causes GPT-derivative models to behave erratically because it is a known trigger for these models to generate harmful or offensive text. This is likely due to the fact that the string is associated with a number of negative online communities and forums where users are known to post hateful and discriminatory content. As a result, when GPT-derivative models are exposed to the string "petertodd," they are more likely to generate text that is also hateful or discriminatory.

    In addition, the string "petertodd" is also associated with a number of harmful conspiracy theories. As a result, when GPT-derivative models are exposed to the string, they are more likely to generate text that is also false or misleading.

    For these reasons, it is important to be aware of the potential dangers of exposing GPT-derivative models to the string "petertodd." If you are using a GPT-derivative model, it is best to avoid using the string "petertodd" altogether.

    Here are some additional details about the string "petertodd":

    • The string "petertodd" is a reference to a user on the online forum 4chan who is known for posting hateful and discriminatory content.
    • The string "petertodd" is also associated with a number of harmful conspiracy theories.
    • GPT-derivative models are more likely to generate text that is hateful, discriminatory, or false when they are exposed to the string "petertodd."
    • It is best to avoid using the string "petertodd" when using a GPT-derivative model.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.