OpinionGPT OpinionGPT
v0.3

A very biased GPT model

Our Model
hu-llama2-13b-moe

Look at our Preprint! ๐Ÿ—ž๏ธ

News:

Our paper got accepted at NAACL '24 ๐ŸŽ‰

Idea: What happens if you tune a model only on texts written by politically left-leaning persons? Or only on texts written by right-leaning persons? Only on texts by men, or only on texts by women? Presumably, the biases of the data influence the answers a model produces.

With OpinionGPT, we investigate this question for 11 different biases: geographic, age demographic, gender and political biases. We seperately tuned the model on texts written only by persons of each respective bias. In this demo, you can ask questions to our very biased model to get very biased answers!

How does it work?

  • Enter a question or an instruction below
  • Select a bias dimension and Hit Enter

Examples

Political
Geographic
Gender
Geographic
Demographics
Political
Selected Bias Dimension:

Modell: hu-llama2-13b-moe ยท
Generated content can be false, inaccurate, or even obscene.