Datasets:
Release validation — SMedDV v5
The commands below were run locally from the repository root. Source data in data/synth/v5/ was read-only. No Hugging Face repo was created or contacted for this release; the live Hub viewer therefore was not tested.
Acceptance checks
| # | Check | Result |
|---|---|---|
| 1 | Row accounting | PASS — 3,527,958 Parquet rows: 3,386,840 train and 141,118 validation. All source labels matched one Parquet row; 0 label/index mismatches. The 155 missing requested indices are all listed in metadata/skips.jsonl. |
| 2 | Split integrity | PASS — the validation indices equal splits/validation_indices.json (141,118 rows); indices are unique, no index crosses splits, and the 3,528,113-index request range is accounted for by rendered rows plus the 155 explicit skips. |
| 3 | Round-trip | PASS — 500 seeded random samples from Parquet shards decoded as WebP, 48 px high; logical and visual labels matched source UTF-8 bytes. 0 sample failures. |
| 4 | Label integrity | PASS — maximum length 64; all 503 non-space dictionary characters are present and no label character is outside the dictionary; minimum occurrences: 800 overall, 755 train, 24 validation. |
| 5 | Shard integrity and checksums | PASS — 19 Parquet shards, 12,092,403,362 bytes; per-shard SHA-256 values and the whole-package checksum verify with no mismatches. Original encoded image bytes accounted for: 11,872,335,840. See data/build_manifest.json and checksums/. |
| 6 | Leakage / prohibited files | PASS — no private absolute paths, credentials, .env, editor/cache files, or loose image files found in the package. |
| 7 | Hugging Face constraints | PASS — 97 package files, max 31 files in a folder, largest file 662,276,687 bytes. Parquet uses the HF Image struct metadata; README frontmatter parses and contains 53 language tags (MRZ is separate). |
| 8 | Reproducibility | PASS — source file hashes, source image content-tree hash, seed, commands, and the source code snapshot commit are recorded in docs/reproducibility.json, docs/source-code.sha256, and data/build_manifest.json. CITATION.cff parses as YAML and its required dataset fields were checked. |
Commands
Parquet build (single process, 2,048-row groups, approximately 650 MB of source image bytes per shard):
PYTHONHASHSEED=0 .venv/bin/python release/v5_hf/tools/build_parquet.py \
--source data/synth/v5 \
--out release/v5_hf/data \
--target-bytes 650000000 \
--row-group-size 2048
Checksum generation and verification:
.venv/bin/python release/v5_hf/tools/write_checksums.py --package release/v5_hf
.venv/bin/python release/v5_hf/tools/write_checksums.py --package release/v5_hf --verify
Full package validation (run from the repository root; the JSON report is written beside v5_hf, not inside the package):
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python release/v5_hf/tools/validate_release.py \
--package release/v5_hf \
--source data/synth/v5 \
--report-json release/v5_validation_results.json
The sample selection seed is 20261031. The Parquet schema carries Hugging Face Image feature metadata and passed the local Arrow smoke test. Hugging Face's dataset-card YAML validation endpoint accepted the README metadata. A live viewer render remains UNVERIFIED until the shards are uploaded.
Remaining review notes
- Wikipedia article-level IDs/URLs were not retained. This package uses dataset/shard-level credit as requested; the sufficiency of aggregate attribution under the upstream CC BY-SA terms remains UNVERIFIED.
- The Noto Sans Hebrew binary's embedded copyright year differs from the downloaded upstream OFL notice; this is disclosed in
ATTRIBUTION.md. The exact matching notice revision remains UNVERIFIED. - No dataset or code was uploaded to Hugging Face.