We release the experiment data dqc-exp-data.pkl to facilitate future research. Each file is a dictionary keyed by (task_name, method_name) tuples. Each entry contains an numpy array of shape (4, 10). The array stores the success rate (averaged over 5 tasks) at 250K, 500K, 750K, 1M training steps for 10 seeds.
See sanity-check.ipynb for a quick example for retrieving the success rate at 1M training steps.