Code for the paper: https://arxiv.org/abs/2306.00740.
All of the main experiments in the paper can be recreated using training/run_training.py
with the appropriate hyperparameters.
The bash scripts under training/
show examples of how to do so; they can be run directly (after modifying training/run_training.py
) if
slurm is available.
The logit analysis (Figure 3 in the paper) that motivated Definition 3.2 can be recreated using training/run_analysis.py
.