# Exact experiment data archive Archive of exact NumPy caches, split metadata and available generator/source files. No re-sampling or train/test relabeling is performed. Split metadata records seeds, tokenizer/source revisions and packing. A training-data evaluation remains a training-data evaluation; independently sampled synthetic tests are distinct from held-out natural-language documents. dataset_manifest.json records checksums; experiment_configs/ records the consuming experiments.