yonghoon96/ogbench-bc-flow-policies

Behaviour-cloned flow policies for OGBench

One BC-pretrained flow policy per OGBench domain: params_300000.pkl, 300k behaviour-cloning steps, seed 10001. Unlike the datasets these are not downloadable from anywhere else, so anything that starts from a pretrained policy on these environments would otherwise have to repeat BC pretraining first.

Each directory keeps the training record beside the checkpoint: flags.json is the full configuration that produced it and eval.csv the evaluation curve during pretraining, so the checkpoint can be verified rather than taken on faith.