omarabb315/baligh-hamasa

Balīgh

Instruction-tuning data for classical Arabic, built from printed books of the Arabic philological tradition. 40,145 records across two books.

• taḥwīl (تحويل) — rewrite modern Arabic, MSA or dialect, into classical Arabic.
• qa — one linguistic fact from the book, asked in a real voice (student, reader, writer, teacher, editor, preacher, learner) and answered in the author's words.
• sharḥ_kāmil — a verse explained under fixed scholarly headings from several of its claims at once.
• ḥiwār — a short dialogue about one verse, in a single persona.
• sharḥ (شرح) — mined verse → commentary pairs.
• gloss — older one-line vocabulary forms, kept for al-Kāmil (Ḥamāsa now uses qa instead).

Books

book	author	records
شرح ديوان الحماسة	al-Marzūqī (d. 421 AH)	32,408
الكامل في اللغة والأدب	al-Mubarrad (d. 285 AH)	7,737

Every record carries meta.book_slug, so any single book can be isolated or excluded.

The central design decision

For the conversion task, the pipeline runs backwards. Rather than asking a model to produce classical Arabic — which would make the training target synthetic — it takes the author's authentic prose as the target and asks Gemini to degrade it into modern MSA and a dialect. Those degraded versions become the user turn.

Every assistant turn in the taḥwīl split is human-written classical Arabic from a printed book. No model wrote the thing the student model learns to imitate.

The same principle governs everything else here. The sharḥ and gloss records are mined, not generated: 17,662 of the 21,802 records are verbatim text lifted out of the books, and each one is checked back against the source before it is kept.

Questions are written for this dataset; answers are the book's. Gemini writes the user turn (the question, the dialect input, the opening of a dialogue). The assistant turn is the author's wording: either a verbatim slice of the book, or a form-only adaptation (supply the subject, drop a dangling «و») that is then scored word-by-word against the entry. 27,284 of 27,785 composed answers are made entirely of the author's words; the rest carry a fidelity score and the list of words that could not be found. Nothing is deleted for a low score — the flag travels with the record.

Q (written for this dataset)  ما المراد بـ (كأس) و(الظنبوب)؟
A (al-Mubarrad's own words)   يقول: لأغيث. وكأس: اسم جارية، وإنما أمرها
                              بإلجام فرسه ليغيث. والظنبوب: مقدم الساق.

404 records are built this way and tagged meta.source = "inverted", and every one of their answers is verified verbatim against the book. A question is rejected if it is too short to be specific, longer than 60% of its answer — a restatement rather than a question — or does not end in a question mark.

How the glosses are extracted

Commentators explain hard words constantly, in prose, with no fixed pattern: الأداهم: القيود, or والوفر: المال الكثير, or a relative clause that never uses a colon at all. A regex over the colon form was tried first and abandoned — it produced pairs like تحقيق الكلام → لم أدر ما يقتضي هذا السؤال, which is a sentence break, not a definition.

Instead Gemini reads the prose and reports what it finds, and every proposal is then verified back against the book: both the term and the definition must occur verbatim in the source, and the text stored is the book's wording, not the model's. A proposal the model invented has nothing to match and is dropped. Across the two books 24,504 proposals were checked and 6,605 unique glosses kept; the rest were rejected or were duplicate terms.

The reverse direction (meaning → word) is only built where the definition identifies one word: at least 12 characters, and not shared with another term in the same book. من أسماء الأسد defines both الليث and الرِّيبال, so it is never asked in reverse.

The context direction shows the word as the poets used it, never as the commentator explained it. عند الحفيظة إن ذو لوثة لانا asks what اللوثة means there, and the answer is القوة. The explaining paragraph is deliberately not used as context: it states the meaning, so it would hand over the answer.

The verse is chosen by the model and then verified, not found by string search. An earlier build searched the book for a line whose letters matched the term, and half the results were homographs from unrelated entries: الدين glossed as recompense attached to a line where it means debt, العنبر glossed as perfume attached to a tribal name, قائم glossed as subduing attached to a sword's hilt. Now each entry's verses and its commentary are read together in one pass, the model names the verse it was glossing, and that claim is checked three ways: the word must occur in the named verse in the verse's own spelling, the verse must belong to the same entry as the gloss, and the verse must not itself state the meaning. 4,074 of 7,589 glosses earn a context this way; the rest ship without one rather than with a wrong one.

Format

Messages format, one JSON object per line:

{
  "id": "hamasa-000100-mined-00",
  "task": "sharh",
  "split": "train",
  "messages": [
    {"role": "system",    "content": "فسِّر ما يلي على طريقة الشُّرَّاح، موضحًا الألفاظ والمعاني."},
    {"role": "user",      "content": "فدت نفسي وما ملكت يميني ... فوارس صدقوا فيهم ظنوني"},
    {"role": "assistant", "content": "لفظه لفظ الخبر. والمعنى معنى الدعاء..."}
  ],
  "meta": {"source": "mined", "book_slug": "hamasa", "entry_id": "hamasa-000100", "...": null}
}

meta carries the same keys on every record so the schema stays uniform and the viewer can filter on it. Keys that don't apply to a task are null.

field	applies to	meaning
book, book_slug, entry_id	all	source book and the entry the record came from
variant	taḥwīl	msa or dialect
dialect	taḥwīl	which dialect, null for the MSA variant
passage_id, granularity	taḥwīl	passage identity; sentence or block
repair_distance	taḥwīl	how far repair moved the text from the raw scan
similarity_to_target	taḥwīl	character overlap with the target — see Limitations
source	sharḥ, gloss	mined (answer lifted from the book) or inverted (answer from the book, question written for this dataset)
miner	sharḥ, gloss	which extractor produced it
poet	sharḥ	attribution line from the book

all.jsonl contains all 40,145 records with the split field, for anyone who prefers to re-split.

Composition

task	records
qa	21,825
sharḥ_kāmil	4,171
ḥiwār	1,789
taḥwīl	3,736
sharḥ	2,546
gloss / gloss_context / gloss_reverse	6,078

split	records
train	36,341
validation	1,744
test	2,060

Splits are assigned per entry, so a verse and its commentary never land on opposite sides of a split.

Conversion pairs cover five dialects alongside MSA: Egyptian (401), Moroccan (377), Levantine (373), Gulf (363), Iraqi (354).

Read this before fine-tuning. qa is now the largest task (21,825). Many of those answers are still short lexical glosses, now with a real question in front. sharḥ_kāmil and ḥiwār teach multi-fact explanations; taḥwīl and mined sharḥ teach sustained classical prose. Filter or re-weight on task, persona, and register. Nothing is lost by dropping a subset.

How it was built

1. Extract the book from MohamedRashad/Al-Maktabah-Al-Shamilah.
2. Stitch page-sized rows back into one continuous text, since rows split mid-sentence. Seams are classified as verse boundary, heading, or mid-sentence, and joined accordingly.
3. Profile the book with Gemini — genre, structural unit, opening formulas, whether it carries a base text — and use that to template later prompts, so a poetry commentary and an adab miscellany are not segmented by the same rules.
4. Segment with Gemini, which returns line numbers only, never text, so a passage can only ever be a verbatim slice of our own file. The hard part is telling an entry-opening verse from a verse cited as evidence inside commentary — roughly half the verses in the Ḥamāsa are citations, and splitting on them destroys the commentary.
5. Repair orthography line by line: OCR damage, spacing, hamza and yāʾ variants, and selective diacritics where they disambiguate.
6. Degrade for taḥwīl; mine verse/commentary pairs and glosses for the rest.
7. Validate and package.

The guards

Two mechanisms do the real work, and both are checks on text rather than instructions to a model.

The repair guard. Repair is the only stage permitted to alter text. An edit is accepted only if the two strings share a canonical form — diacritics stripped, letter variants folded — and every difference falls in an explicit allowlist (diacritic, letter_fold, whitespace, punctuation, duplicate_letter, archaic_variant). Anything that changes, adds, removes or reorders a word is rejected and the book's own wording stands. All 23,999 accepted repairs re-verify clean under an independent second pass.

The verbatim guard. Everything mined must be found in the book. Matching is done on the consonantal skeleton, so the model is not punished for dropping a diacritic, but the text kept is the source's. This guard is also what makes model error visible: it initially rejected 843 sound glosses because a term and its definition sat on opposite sides of a line break, which is page layout rather than a sentence boundary — the check now reads a line together with its neighbours.

Verification

Every check runs in validate.py and passes on this release, per book:

check	Ḥamāsa	al-Kāmil
entries tile the book with no gaps or overlaps	885 / 10,778 lines	798 / 14,164 lines
passages are verbatim slices of the source	3,174 / 3,174	1,057 / 1,057
no verse cut mid-bayt	3,174 / 3,174	1,057 / 1,057
no mid-word boundaries	3,174 / 3,174	1,057 / 1,057
repair guard re-verified independently	9,913 / 9,913	14,086 / 14,086
taḥwīl targets byte-identical to source	2,462 / 2,462	1,274 / 1,274
mined pairs found verbatim in the book	12,522 / 12,522	6,326 / 6,326
no context contains its own answer	2,665 / 2,665	1,410 / 1,410
inverted answers found verbatim in the book	19 / 19	385 / 385
record ids unique	14,099 / 14,099	7,737 / 7,737
gloss context belongs to its own entry	2,665 / 2,665	1,410 / 1,410

Limitations

Read these before training on it.

• The gloss forms dominate and are terse. See the warning under Composition.
• No semantic-fidelity check on taḥwīl. similarity_to_target is fuzz.ratio over the consonantal skeleton — pure character overlap. It is used to reject rewrites above 90% similarity, meaning the model failed to modernise. It cannot detect the opposite failure, where a rewrite drifts off the original meaning. Nothing in the pipeline verifies that the modern and classical sides say the same thing.
• Dialect authenticity is unverified. The dialect turns are Gemini's rendering, not native-speaker text, and no speaker of any of the five dialects has reviewed them.
• A gloss is verified as present, not as correctly paired. The guard proves both halves are the book's words near each other; it cannot prove the author meant the second as a definition of the first. Spot-checking found sound pairs, but no exhaustive review was done.
• Sense drift survives inside an entry. Restricting a context to the gloss's own entry removes homographs from elsewhere in the book but not from the same passage. المازن is glossed بيض النمل (ants' eggs) and attached to لو كنت من مازن, where it is the tribe — the commentator's lexical aside and the poet's usage are different words sharing an entry. Rate unmeasured.
• Reverse glosses use a shallow uniqueness rule. A definition qualifies if it is 12+ characters and not an exact duplicate within its book. Near synonyms therefore both survive: كل هارب من شيء → أجفل عنه and كل هارب من شيء مسرع → مجفلٌ وجافلٌ are both present, and the question is ambiguous between them. 80 definitions are a substring of another definition.
• 52 passages carry an unclosed quotation mark inherited from the source scan. The damage is in the printed text, not introduced here.
• al-Kāmil's sharḥ split is inverted, not mined. It is an adab miscellany with no regular verse-then-commentary structure, so no explanation miner applies. Its 385 sharḥ records pair al-Mubarrad's own paragraph with a question written for this dataset (meta.source = "inverted"). The answers are his; the questions are not. 19 Ḥamāsa records are the same.
• The questions in inverted records are unreviewed. They are checked for shape — specific enough, not a restatement, actually interrogative — but nothing verifies that a question is the best one for its paragraph, or that it does not presuppose something the paragraph never says.
• Segmentation is not bit-reproducible. Two runs over identical cached model responses produced 3,171 and 3,174 spans, a 0.09% difference traced to iteration order in window reconciliation.
• Most taḥwīl passages are not self-contained. These are commentaries, so much prose refers back to a verse (يقول، ومعنى البيت). Such passages still yield sound pairs — both sides refer to the same absent verse — but the judgement is recorded so the mix can be weighted.
• Two books, two authors, two genres. A poetry commentary and an adab miscellany, both from the philological tradition. Expect their idiom and their concerns.

Source

Both books are long out of copyright; al-Mubarrad died in 285 AH / 898 CE and al-Marzūqī in 421 AH / 1030 CE. The digital text comes from Maktabat al-Shāmila via MohamedRashad/Al-Maktabah-Al-Shamilah and carries that edition's OCR characteristics.

Citation

@misc{baligh_2026,
  title  = {Bal{\=i}gh: classical Arabic instruction data},
  author = {omarabb315},
  year   = {2026},
  url    = {https://huggingface.co/datasets/omarabb315/baligh-hamasa}
}