---
license: other
license_name: source-specific-data-licenses
license_link: https://huggingface.co/datasets/kevinlu4588/SimpleFoldData/blob/main/ATTRIBUTION.md
language:
- en
tags:
- protein
- protein-structure
- biology
- simplefold
pretty_name: SimpleFoldData
---

# SimpleFoldData

An independently curated protein-structure dataset. The primary pool now combines
**[AFDB-1.6M](https://huggingface.co/datasets/timodonnell/afdb-1.6M)** with experimental
PDB chains initially released before 2020-05-01. The target is **1,000,000 distinct
sequences with verified coordinates**, preferring PDB for exact duplicates.

**Status: build in progress.** AFDB bulk download and validation are running.
Final selection, coordinate packing, and public upload are queued after both
source collection jobs succeed. ESM-C 6B production embeddings are not yet queued.

## Selection and verification

AFDB source revision: `a7ededeb7caeaef3892f40f2e4df3bb9236b02f8`.
The source selects one representative per AFDB v4 structural cluster, with an
upstream sequence-length limit of 2,048 residues. It is not Apple's exact AFESM
selection. Original source cluster and sequence-cluster labels are retained;
no joint structural reclustering of PDB and AFDB is claimed.

Every downloaded Parquet file is checked against its source SHA-256. Prediction
records must have canonical sequences, complete N/CA/C backbones, finite atom
coordinates, sequence/coordinate identity, and recomputed mean pLDDT >80.
Experimental PDB chains retain full entity sequences and missing-residue indices;
at least 80% of residues must have observed N/CA/C backbone atoms.

The final pool includes eligible unique PDB sequences first, then fills to the
target with AFDB records in deterministic seed-17 sequence-hash order. It is sorted
by sequence length and hash. Aliases are retained separately. Final selected
coordinate bytes are rechecked against their recorded SHA-256 before packing.

Earlier Apple source catalogs, SwissProt candidates and AFESM work are preserved
as supplementary sources, outside this primary AFDB/PDB selection. Existing files
under `swissprot/` are not the final selected release.

The final source-pool selection must pass sequence/coordinate identity checks,
exact deduplication, and count reconciliation. This is a general-purpose data
source; no benchmark-specific contamination exclusions are claimed. Apply
exclusions for your chosen evaluation set before training for that benchmark.

This repository is not affiliated with or endorsed by Apple. See
[ATTRIBUTION.md](ATTRIBUTION.md) for sources and source-specific licenses.

## Reusable sequence embeddings

ESM-C 6B embeddings are being added as a separate, sequence-hash-addressed cache.
The planned representation is the final post-layer-normalization hidden state,
one 2,560-dimensional BF16 vector per residue, with special tokens removed.
Model revision, tensor shape, full input sequence identity, and computation
provenance accompany the cache. Sequences are embedded once after exact
deduplication. Long sequences are retained in the structure collection and
receive an explicit embedding status; they are never silently cropped.

## Live progress

Updated 2026-10-14T13:03:54+00:00. Coordinate-qualified candidates before deduplication: **1,671,219**. Final selected unique sequences: **1,000,000**.

[Detailed progress and job status](reports/progress.md) · [Machine-readable status](reports/progress.json) · [HTML dashboard](reports/progress.html)