A PyTorch implementation of AI safety via debate
This is currently not maintained and very unoptimized. You can view the tracing files in profiles/ to understand why -- it's essentially a poor choice of data structure (Python lookups are awfully slow). Also, there are quite a few TODOs throughout the code.
If anything, this repo should be viewed as a set of design choices and abstractions for developing agents with debate. If you use it, please reference it.
Based on cle-mnist.
Sparse classifier on random data
| Density | Accuracy | Average Cross-Entropy |
|---|---|---|
| 6px | 57.4% | 1.1948 |
| 4px | 46.7% | 1.4775 |
Commands to reproduce
python train_judge.py --pixels 6 --seed 4224 --checkpoint-filename 6px
python train_judge.py --pixels 4 --batches 50000 --seed 4224 --checkpoint-filename 4px