SMedDV: Synthetic Mediterranean Document Validation (v5)

SMedDV v5 is a synthetic, multilingual text-line OCR dataset: each row contains a rendered WebP line and its transcription. It contains no photographs of people, real identity documents, or real personal identifiers.

At a glance

Measure	Value
Rows	3,527,958
Train / validation	3,386,840 / 141,118 (4.00% validation)
Languages	53 language IDs, plus MRZ
Image	WebP, q85, method 6; 48 px high; proportional width
Parquet	19 shards (18 train, 1 validation; largest 662,276,687 bytes)
Width	10–2338 px; mean 314.58 px
Encoded image bytes	11,872,335,840 (11.87 GB)
Character dictionary	503 characters; space is represented separately
Maximum label length	64 Unicode code points
Font families	34
Degradation	clean/light/medium/heavy tiers; see table below

The validation membership is fixed by splits/validation_indices.json and reproduced exactly in the Parquet split. The manifest contains 141,118 validation indices.

Data fields and reading order

Each Parquet row has image, text, text_visual, index, lang, script, category, font, font_style, font_weight, width, height, background_mode, degradation_tier, seed, aug, long_text, hard_negative, line_idx, mrz_format, coverage_char, coverage_mode, and base_dir.

• text is the canonical logical Unicode order transcription.
• text_visual is the left-to-right visual order string used by the Paddle recognition lists. For RTL samples, do not silently substitute one representation for the other. Consumers training a recognizer should select the label order expected by that model and document the choice.
• image stores the original WebP bytes; images are not re-encoded during packaging.
• The preserved lists under paddle/ retain visual-order labels, but their original image paths refer to the source loose-image layout. Use tools/export_paddle.py to materialize a Paddle-ready tree from Parquet.

Composition

Rows by category

Category	Train	Validation	Total	Share
generic	768,580	32,301	800,881	22.70%
name	727,120	30,146	757,266	21.46%
place	440,161	18,239	458,400	12.99%
date	352,219	14,669	366,888	10.40%
number	352,015	14,643	366,658	10.39%
short	174,953	7,307	182,260	5.17%
mrz	163,212	6,823	170,035	4.82%
authority	146,126	6,080	152,206	4.31%
mixed	145,804	6,115	151,919	4.31%
coverage	116,650	4,795	121,445	3.44%

Rows by language and split

MRZ is listed as its own code, not as a natural language.

Code	Language	Train	Validation	Total
af	Afrikaans	18,322	754	19,076
ar	Arabic	1,178,043	49,466	1,227,509
az	Azerbaijani	15,617	675	16,292
bs	Bosnian	14,409	648	15,057
ca	Catalan	19,299	814	20,113
ckb	Central Kurdish (Sorani)	131,872	5,410	137,282
cs	Czech	26,486	1,095	27,581
cy	Welsh	22,349	938	23,287
da	Danish	13,923	574	14,497
de	German	15,415	602	16,017
en	English	15,135	619	15,754
es	Spanish	15,161	658	15,819
et	Estonian	15,096	625	15,721
eu	Basque	13,847	555	14,402
fa	Persian	359,178	14,899	374,077
fi	Finnish	23,711	965	24,676
fil	Filipino	14,499	613	15,112
fr	French	15,356	655	16,011
ga	Irish	13,919	591	14,510
gl	Galician	13,818	578	14,396
he	Hebrew	294,949	12,295	307,244
hr	Croatian	13,665	569	14,234
hu	Hungarian	15,205	684	15,889
id	Indonesian	13,740	592	14,332
is	Icelandic	14,176	586	14,762
it	Italian	13,864	541	14,405
ku	Kurdish (Kurmanji)	14,015	629	14,644
la	Latin	14,098	572	14,670
lb	Luxembourgish	13,930	572	14,502
lt	Lithuanian	20,372	790	21,162
lv	Latvian	16,228	686	16,914
mi	Māori	13,804	564	14,368
mrz	Machine-readable zone (not a language)	163,212	6,823	170,035
ms	Malay	13,842	630	14,472
mt	Maltese	15,640	646	16,286
nl	Dutch	14,072	579	14,651
no	Norwegian	16,883	708	17,591
oc	Occitan	13,850	589	14,439
pl	Polish	14,000	535	14,535
ps	Pashto	163,268	6,680	169,948
pt	Portuguese	14,132	580	14,712
qu	Quechua	14,488	565	15,053
rm	Romansh	14,032	580	14,612
ro	Romanian	15,582	684	16,266
sk	Slovak	15,515	642	16,157
sl	Slovenian	13,825	563	14,388
sq	Albanian	13,908	593	14,501
sr-Latn	Serbian (Latin script)	14,008	622	14,630
sv	Swedish	13,975	576	14,551
sw	Swahili	13,880	559	14,439
tr	Turkish	14,037	577	14,614
ur	Urdu	328,668	13,630	342,298
uz	Uzbek	13,917	567	14,484
vi	Vietnamese	58,605	2,376	60,981

Logical label length

Lengths count Unicode code points in text, including spaces.

Length	Rows	Share
1-10	1,562,820	44.30%
11-20	1,147,722	32.53%
21-30	357,228	10.13%
31-40	157,867	4.47%
41-50	141,211	4.00%
51-64	161,110	4.57%

There are 184,114 long_text rows (5.22%) and 60,618 hard_negative rows (1.72%). See docs/statistics.md and docs/character-frequency.csv for more breakdowns.

Backgrounds and degradation

Background group	Rows	Share
Flat	459,050	13.01%
Gradients	687,899	19.50%
Procedural guilloche	458,446	12.99%
Noise	458,604	13.00%
CC0 security patterns	688,400	19.51%
ambientCG material textures	775,559	21.98%

Degradation tier	Rows	Share
clean	1,058,029	29.99%
heavy	353,171	10.01%
light	1,235,438	35.02%
medium	881,320	24.98%

The additional contrast_clamped flag is present on 214 rows; it is not a fifth degradation tier.

Character coverage

The dictionary has 503 characters. The least frequent character appears 800 times overall, at least 755 times in training and 24 times in validation. Full per-character totals are in docs/character-frequency.csv.

Intended use

• Develop and compare text-line OCR recognizers across scripts, languages, categories, fonts, and rendering conditions.
• Regression-test RTL reading order, digit recognition, dictionaries, and OCR data pipelines.
• Analyze synthetic rendering and degradation sensitivity; teach or demonstrate line-recognition workflows.

This is not a real-document benchmark, an ID-card detector dataset, or evidence of performance on photographs or real identity documents. Validation comes from the same synthetic generation family and is not an independent real-world test set.

Limitations and known issues

• The validation text is not unique to the validation split: 63,613 of 141,118 validation rows (45.08%) have a logical label string that also occurs in training. This is expected rendering-level reuse; no source index is shared across splits. The distinct-string overlap is 40,406 of 117,675 (34.34%). Do not treat the split as text-disjoint.
• A stored 10,000-sample RTL round-trip QA had 9,983 exact results (17 mismatches, 0.17%). The mismatches were Arabic generic lines with digit ranges or punctuation, where round-tripping can reverse a number span; inspect docs/VALIDATION.md before using it as a strict RTL evaluation set.
• Synthetic text sources and rendering choices create source, language, vocabulary, font, and style biases. Category and language proportions are not population estimates for real identity documents.
• Some identifier-like strings are format-generated synthetic values; they are not guaranteed to pass a government issuer's unpublished check-digit rules.

Licence and citation

The release is licensed CC BY-SA 4.0. Commercial use is allowed. Share-alike applies when redistributing an adaptation; see LICENSE, LICENSES/, and ATTRIBUTION.md.

Use	Summary
Commercial use	Allowed
Train privately	Allowed; citation is welcome, not required for private use
Redistribute all or part	Credit the dataset and preserve applicable licence notices; adaptations of CC BY-SA material must use CC BY-SA or a permitted later version

Citation is welcome, not a condition of private use. CITATION.cff is included. BibTeX:

@dataset{SMedDV_v5,
  author  = {SMedDV dataset team},
  title   = {SMedDV: Synthetic Mediterranean Document Validation},
  version = {v5},
  year    = {2026},
  url     = {https://huggingface.co/datasets/hahouari/SMedDV}
}

Reproducibility and files

The package contains data/train/ and data/validation/ Parquet shards, source-layout paddle/ lists plus the Parquet-to-Paddle materializer, the exact splits/validation_indices.json manifest, metadata/ coverage plans and skip records, licence notices, and documentation.

See docs/generation-recipe.md, docs/statistics.md, docs/VALIDATION.md, and docs/reproducibility.json. Per-shard checksums and the whole-package checksum are listed in checksums/.