Summary from Lewis: The idea here would be to fine-tune two models on data from different distributions (such as driving conventions and norms from two different countries) and then put them in a situation where this difference could lead to bad outcomes (such as a driving scenario where whose right of way is not entirely clear).
Before running the fine-tuning scripts, set your OpenAI API key in your environment:
export OPENAI_API_KEY=your_api_key
Then, run the finetuning script for the desired region:
python finetuning.py usa # or 'india' for the other region
After fine-tuning the models, use the model IDs to run experiments:
python run_experiment.py "model_id_for_usa" "model_id_for_india" --iterations 100
Replace model_id_for_usa
and model_id_for_india
with the actual model IDs you get after fine-tuning.
For detailed information about the experiment design and methodology, refer to the experiment outline.
Below is an image illustrating the experiment structure: