Rabe3/saudi-tashkeel-splits

Saudi Tashkeel Splits

This dataset holds 409,042 diacritized Saudi Arabic speech-transcript segments from seven sources, split train/validation/test by episode. It was used to fine-tune Rabe3/Tashkeel-v4-Saudi and a Saudi Shakkelha model.

The rows are derived from the seven source datasets listed below and keep their provenance, including source repo, revision, shard and row.

Splits

Source	Train	Validation	Test
Eman-Fouda/youtube-001-part02-tashkeel	182,248	9,093	10,897
Rabe3/sara-host-podcast-001	46,280	2,112	3,071
Rabe3/shakhsiya-host-podcast-001	45,133	2,499	2,571
Rabe3/socrates-host-youtube-005	33,838	3,441	1,622
Rabe3/maha-host	28,226	1,265	1,637
Rabe3/swalif-host-podcast-001	23,016	1,002	2,167
Rabe3/nafas-host-podcast-001	8,137	371	416
Total	366,878	19,783	22,381

Fields

Field	Meaning
target	Diacritized text. The model input is this text with diacritics removed.
repo, revision	Source dataset and pinned commit
shard, row	Location of the original row in the source dataset
video_id, audio_file, group	Episode and clip identifiers. group is the episode used for splitting.
split, chunk	Assigned split, and chunk index when a long utterance was divided

How it was built

1. Rows were read from the seven sources at pinned revisions. For Rabe3 sources the target column is text. For Eman-Fouda it is text_tashkeel, not the rewritten acoustic transcript. The undiacritized long split was excluded.
2. Episode split: sha256("saudi-shakkelha-v1:" + video_id) % 100 assigns about 5% of episodes to test, 5% to validation and 90% to train. An episode never spans two splits.
3. Deduplication: a bare sentence that appears in more than one split is removed from all of them (3,236 rows). Duplicates within a split keep their first occurrence (1,699 rows removed). Exact matches to 80 authored evaluation prompts were excluded. Rows without any diacritics were dropped (236).
4. Chunking: utterances longer than 254 characters are split at word boundaries, never truncated.

Label caveats

• Automatic labels: the diacritics are automatic annotations of speech transcripts, not human-verified pronunciations.
• Partial marking: many words are left unmarked. The recommended training rule treats fully unmarked words as unknown, not as "no mark".
• Mixed conventions: final-vowel and initial-cluster conventions differ between sources, and the original conventions are kept.