Feature Reconstruction
This source-first release excludes all precomputed embeddings, graph tensors, structural features, and packed feature caches. Reconstruct derived features from data/proteins.parquet using the published PhaseFlow preprocessing workflow and its documented pretrained components. Store all generated artifacts outside this dataset directory so the distributed inventory remains stable.
Use protein_id and sequence_sha256 together to bind reconstructed features to source records. Confirm sequence lengths and hashes before reusing any cache. data/base_training_schedule.parquet and data/region_supervision.parquet are protocol/supervision tables, not feature caches.
The files under benchmark/ are evaluation inputs. Feature generation for benchmark sequences may be necessary for inference, but the benchmark must remain outside training, hyperparameter selection, and checkpoint selection.