Materialized datasets (HF Arrow) that the DUO/looped-diffusion trainer loads
directly via datasets.load_from_disk(...). Two splits:
openwebtext-train_train_bs1024_wrapped.dat/ — train
openwebtext-valid_validation_bs1024_wrapped.dat/ — validation
Recipe: openwebtext -> GPT-2 tokenizer -> concatenate + chunk into blocks of
1022 tokens, each wrapped with [BOS] ... [EOS] (block size 1024). This is the
EXACT packing… See the full description on the dataset page:
https://huggingface.co/datasets/saitejautpala/openwebtext-gpt2-packed-1024.
