Data Dictionary
Training data
| File | Unit of observation | Required keys | Description |
|---|---|---|---|
data/proteins.parquet |
protein sequence | protein_id, sequence_sha256 |
Canonical training protein sequences, sequence lengths, and accession fields. |
data/training_units.parquet |
protein training record | protein_id, sequence_sha256 |
LLPS labels, weights, evidence roles, source/protocol metadata, and teacher-derived scalar covariates. |
data/training_plan/batch_plan_epoch_*.parquet |
planned batch placement | dataset_index, plan_dataset_index, protein_id, sequence_sha256 |
Fixed three-epoch 8-GPU batch protocol. |
Join training_units.parquet to proteins.parquet on both protein_id and sequence_sha256. A training-plan row must resolve by dataset_index and plan_dataset_index to the same protein_id and sequence_sha256 in training_units.parquet; do not treat plan indices as stable external identifiers.
Evaluation benchmark
| File | Unit of observation | Description |
|---|---|---|
benchmark/proteins.csv |
benchmark protein | Evaluation sequences, labels, conditions, and source text. |
benchmark/constructs.csv |
assay construct | Experimental construct sequences, labels, conditions, and source text. The first two rows encode the original header representation; the release copy is UTF-8 after a lossless GB18030-to-UTF-8 transcode. |
benchmark/panel_membership.csv |
panel-protein membership | Named evaluation panels, labels, classes, dataset tokens, and descriptions. |
The entire benchmark/ directory is evaluation-only. Do not include its records in fitting, threshold optimization, or checkpoint selection for benchmark reporting.
Configuration and integrity metadata
configs/dataset_contract.yaml and configs/training.yaml document the public package contract and training settings. metadata/feature_reproduction.md defines the source-first feature policy. metadata/file_inventory.csv lists every package file other than itself with byte size and SHA-256 digest.