Note: This code is a work in progress. It will change, and hopefully get better, during the next few months.
This repository provides code to reproduce the experiments from AI Safety via Debate (blogpost).
On top of that we run additional experiment on MNIST as well as FashionMNIST data and train classifiers from debate results.
Install the python dependencies by running the following in a python 3.6
pip install -r requirements.txt
All code is located in the ai-safety-debate folder.
- To train a judge use
train_judge.py - To run individual debates use
run_debate.py - To evaluate the accuracy of a judge combined with debate use
amplify_judge_with_debate.py - To use debate to train a classifier use
train_classifier_via_debate.py
We use sacred for tracking experiments. The results are typically stored in the experiments and amplification_experiments folders. Scripts that use sacred have parameters specified in a config function. To specify values for these parameters, use the with statement, e.g.
python ai-safety-debate/run_debate.py with judge_path=ai-safety-debate/saved_models/mnist4