Document one-million-structure source pool and curation scope

ATTRIBUTION.md

@@ -0,0 +1,23 @@
+# Source attribution
+
+- SimpleFold: Wang et al., *SimpleFold: Folding Proteins is Simpler than You Think*.
+  [Paper](https://arxiv.org/abs/2509.18480),
+  [official code and target lists](https://github.com/apple-aiml-research/ml-simplefold).
+  Author target lists are identified by retrieval URL and SHA-256 in the source
+  ledger. Apple's software MIT license is not asserted to replace data licenses.
+- AFESM: Yeo et al., *Metagenomic-scale analysis of the predicted protein structure universe*.
+  [Cluster resource](https://afesm.foldseek.com/),
+  [analysis repository](https://github.com/steineggerlab/afesm-analysis).
+- AlphaFold Protein Structure Database: Jumper et al. and Varadi et al.
+  [AFDB](https://alphafold.ebi.ac.uk/). Downloaded AFDB structure headers specify
+  CC BY 4.0. Keep original attribution with redistributed coordinates.
+- ESM Metagenomic Atlas: Lin et al., *Evolutionary-scale prediction of atomic-level
+  protein structure with a language model*.
+  [Source and license](https://github.com/facebookresearch/esm#license).
+  Atlas data are CC BY 4.0; copyright Meta Platforms, Inc.
+- Protein Data Bank: [wwPDB](https://www.wwpdb.org/).
+  Each derived chain record will retain its PDB accession, chain identity,
+  initial release date, source locator, and coordinate hash.
+
+The repository uses source-specific data licenses. Dataset preparation does not
+transfer authorship of the original structures or predictions.

README.md

@@ -0,0 +1,49 @@
+---
+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 prepared, versioned protein-structure dataset based on the
+public target lists used by Apple's SimpleFold. The intended first release is
+**1,000,000 distinct protein sequences with validated structures**, combining
+PDB, author-listed AFDB SwissProt, and cluster-balanced AFESM representatives.
+
+**Status: build in progress. No training-ready release has been declared.**
+The initial catalog records published IDs and structural-cluster membership.
+Catalog membership does not imply that a sequence has been recovered, a
+structure verified, or benchmark homology exclusions completed.
+
+## Selection and verification
+
+The build retains Apple's full source inventory and a deterministic seed-17
+cluster ordering. It will deduplicate exact sequences while keeping all source
+aliases, source versions, structure hashes, and historical cluster assignments.
+PDB candidates use an initial-release cutoff before 2020-05-01. SwissProt
+confidence filters are mean pLDDT >85 and population standard deviation <15;
+AFESM confidence must exceed 80 on the 0–100 scale. No 256/1024-residue admission
+cap is imposed. Missing identities, changed versions, inaccessible structures,
+and unresolved fragments remain explicit in audit ledgers.
+
+Author lists refer to historical predictions. Retrieved modern payloads are
+identified by their actual source version and hash; they are not represented
+as original historical bytes. Structural-cluster labels initially reproduce
+the author's membership, not a newly recomputed clustering of changed models.
+
+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.