BalDataset

This public collection includes the completed redistributable prose, lexical, translation-candidate, annotation and language-review exports. Seven image-only educational EPUBs and original annotation files are in resources/. The Academy collection remains in the owner-private companion.

No project native-speaker verification or production-quality certification has occurred. There are zero verified chat/instruction examples. Review-only material is not approved training text. No new Balochi text was generated by this export; upstream authorship claims remain unverified where marked.

Configurations and actual counts

Configuration	Records
prose_candidates	293
recommended_prose	17
asjp_wordlist	103
gatitos_lexicon	3,994
translation_candidates	157
localization_gpl_reference	3,906
ud_reference	300
language_review_only	5,073

The default prose_candidates configuration contains 43,273 whitespace word occurrences across its train/validation/test partitions. recommended_prose is an overlapping source/integrity-prioritized view, not an independent corpus. Never add the two views or their JSONL mirrors. Translation candidates can overlap prose sentences; lexical entries and annotations are different units. Language-review rows preserve language=und where uncertain and are outside prose training. Quality scores are uncalibrated screening heuristics.

The public BalChat scratch model used only the 16-record recommended commercial training split (15,818 words). This larger release was not all used to train that model. BalChat model.

Loading

from datasets import load_dataset
data = load_dataset("shah-bakhsh/BalDataset", "prose_candidates")
recommended = load_dataset("shah-bakhsh/BalDataset", "recommended_prose")

Select reference/review configurations explicitly and inspect record_json and their source terms. The fixed prose train/validation/test grouping is preserved. Auxiliary reference splits do not establish an independent generative benchmark. JSONL mirrors are downloads of the same records, not extra examples. Empty splits are not registered in the viewer; the public recommended view has no test records.

Rights and attribution

The collection is not a single permissive-license dataset. Each record retains its source_license, source_url, attribution/history and original record. CC BY attribution, CC BY-SA share-alike and GPL localization conditions apply to the corresponding material. GPL software-localization templates remain separate from CC prose; original editable pinned localization JSON and COPYING are included. See LICENSES.md, sources.json and licenses/.

Academy material is private noncommercial research only, with CC BY-NC-SA/source-access restrictions retained. Public rehosting is not certified. Dictionary headword/meaning alignment is unknown; sentence-bank authorship and two translated novels' underlying rights need review. Those containers remain outside default prose training. No broader data/model/output rights conclusion is made.

Processing and limitations

Conservative existing cleanup preserves Arabic/Latin orthography, diacritics and joining controls. Original records and hashes remain in record_json. Existing contact-pattern exclusions are retained; full personal-data absence is not certified. Source manifests were checked before export. Image-only book pages have no OCR/native proofreading and are not counted as story text. See statistics.json and provenance.json; manifest.json records every uploaded payload file's SHA256. Sources with unknown redistribution rights and privacy-rejected text are excluded.

Balochi OCR image/text pairs (BalOCR v2)

The separate complete OCR download contains 11,195 retained image/text pairs: 10,265 real Arabic-script pairs and 930 explicitly synthetic printed Latin images. The Latin images render 465 spans from 11 licensed original documents in two OFL fonts; they are not real Latin scans or additional authored language documents.

The corrected v2 experiment has 11,043 training examples and 48 diagnostics. Its Arabic diagnostics are not verified unseen-book examples; the Latin diagnostics are synthetic. The original Arabic archive declares CC BY 4.0 independently of the private noncommercial Academy text collection. Source attribution, exact Latin text snapshots, fonts/licenses, original metadata and quarantined records accompany the complete package. Project native-speaker reviews: zero.

Use the verified downloader with --complete for every canonical pair, source metadata and quarantine material. Without --complete, it downloads the smaller corrected training experiment. The complete package is 491,991,763 bytes; the independent Linux download matched SHA256 c1ecdbfb0e1f09c7058d7d4c25a43d1e5cbc311107cbbc6e81a0d9286717c26a. These image examples are separate from the text configurations above and must not be added to their document counts.

BalOCR code and SDK include the frozen split configuration, training commands and validation reports. Full BalOCR model training remains incomplete because CUDA currently cannot initialize on the available server. No final trained BalOCR weights or production accuracy are claimed. The superseded v1 diagnostic split is retained only as historical evidence and an image carrier.