asfar / README.md
HeshamHaroon's picture
docs: README.md — complete 121-title inventory, reclassify tta### as pdf_code
156ba3a verified
|
Raw
History Blame Contribute Delete
17.4 kB
---
language:
- ar
license: cc-by-4.0
pretty_name: Asfar Arabic Heritage Text Corpus
size_categories:
- 100K<n<1M
task_categories:
- text-generation
- fill-mask
- feature-extraction
tags:
- arabic
- classical-arabic
- heritage
- turath
- language-modeling
- corpus
configs:
- config_name: default
data_files:
- split: train
path: data/train-*.parquet
dataset_info:
features:
- name: item_id
dtype: string
- name: collection
dtype: string
- name: book_title
dtype: string
- name: book_title_source
dtype: string
- name: book_raw
dtype: string
- name: pdf_filename
dtype: string
- name: page
dtype: int64
- name: total_pages
dtype: int64
- name: text
dtype: string
- name: raw_text
dtype: string
- name: text_norm
dtype: string
- name: char_count
dtype: int64
- name: char_count_norm
dtype: int64
- name: arabic_ratio
dtype: float64
- name: arabic_ratio_norm
dtype: float64
- name: di_confidence
dtype: float64
- name: pdf_url
dtype: string
- name: source_detail_url
dtype: string
- name: shard
dtype: string
splits:
- name: train
num_examples: 123062
---
# Asfar — أَسْفَار
**A page-level corpus of classical Arabic heritage.**
123,062 pages · 461 PDF volumes · 6 Archive.org collections · Modern & Classical Arabic.
*`Asfar` (أَسْفَار, plural of سِفْر — "tome, large book") is a cleaned, page-indexed corpus of Arabic heritage literature — grammar, lexicography, adab, history, and poetry commentary — drawn from public-domain scans on Archive.org and normalized for language-model training.*
---
## Summary
| | |
|---|---|
| Rows (pages) | **123,062** |
| Unique PDF volumes | **461** |
| Unique parsed book titles | **121** (the other 339 PDFs were uploaded under filename codes — see below) |
| Archive.org collections | **6** |
| Language | Arabic (Modern Standard and Classical) |
| Median chars per page | 1,245 |
| Median Arabic-letter ratio | 0.971 |
| Split | single `train` split |
| Format | Parquet, 5 shards, zstd-compressed |
| License | CC-BY-4.0 (metadata & normalization work); underlying books are public-domain — cite the book |
---
## Quickstart
```python
from datasets import load_dataset
ds = load_dataset("HeshamHaroon/asfar", split="train")
print(ds)
print(ds[0]["text_norm"][:200])
```
To use only rows with a recovered book title:
```python
titled = ds.filter(lambda r: r["book_title_source"] == "parsed_from_pdf_name")
# ≈ 45,013 pages across 121 named books
```
---
## Columns
### Text
| Column | Dtype | Purpose |
|---|---|---|
| **`text_norm`** | `string` | Primary training column. NFC-normalized, tatweel removed, tashkeel stripped, lam-alef ligatures expanded, whitespace collapsed. |
| `text` | `string` | OCR output after an LLM cleaning pass (publisher-supplied). Keep diacritics if present. |
| `raw_text` | `string` | Pre-cleaning OCR. Useful as the noisy side of a `raw_text` → `text` OCR-denoising pair. |
### Source identity
| Column | Dtype | Purpose |
|---|---|---|
| **`book_title`** | `string` | Cleaned book title. For PDFs whose filename encoded the title (e.g. `" 0 1 1 خزانة الأدب.pdf"`), leading volume/part/issue numeric markers are stripped to yield `"خزانة الأدب ولب لباب لسان العرب"`. For PDFs whose filename is a bare code (e.g. `"10668.pdf"`, `"a025.pdf"`, `"tta0104.pdf"`), `book_title` equals the code — the actual title is not present in upstream metadata. Inspect `book_title_source` to tell the two cases apart. |
| **`book_title_source`** | `string` | One of: `parsed_from_pdf_name` (45,013 rows / 122 PDFs / 121 distinct titles), `pdf_code` (78,049 rows / 339 PDFs), `pdf_stem` (edge cases). |
| `book_raw` | `string` | The original, unprocessed `book` string from the upstream dataset (= PDF filename stem). Kept for auditing the parsing. |
| `collection` | `string` | The **Archive.org collection** (6 distinct) this PDF was harvested from — a coarse grouping over many books, not a book title. Renamed from the upstream dataset's `item_title` column, which had collection strings under a misleading name. |
| `item_id` | `string` | Archive.org item identifier (links to `source_detail_url`). |
| `pdf_filename` | `string` | Original PDF filename as uploaded to Archive.org. |
| `pdf_url` | `string` | Direct download URL for the source PDF on Archive.org. |
| `source_detail_url` | `string` | Archive.org item/collection detail page. |
| `shard` | `string` | Original `az_NNNN` config name from the upstream `HeshamHaroon/arabic-turath-ocr` dataset. Provenance only. |
### Position
| Column | Dtype | Purpose |
|---|---|---|
| `page` | `int64` | Page index within the PDF. |
| `total_pages` | `int64` | Total pages in the PDF. |
### Quality signals
| Column | Dtype | Purpose |
|---|---|---|
| `char_count`, `char_count_norm` | `int64` | Character count of `text` and `text_norm`. |
| `arabic_ratio`, `arabic_ratio_norm` | `float64` | Fraction of non-whitespace characters in the Arabic Unicode block. |
| `di_confidence` | `float64` | Document-intelligence confidence reported by the upstream OCR pipeline. |
---
## Archive.org collections
Six collections account for all 461 volumes. Three of them use codes as PDF filenames — the titles aren't recoverable from metadata alone. The other three use readable Arabic titles.
| pages | PDFs | collection | filename convention |
|---:|---:|---|---|
| 45,249 | 229 | كتب عربية - طبعات قديمة | numeric codes (e.g. `10668.pdf`, `30283.pdf`) |
| 29,450 | 100 | كتب منوعة - المائة الثانية pdf | `a###` codes (e.g. `a025.pdf`, `a077.pdf`) |
| 28,202 | 99 | كتب متنوعة في اللغة العربية ولسانها | Arabic titles |
| 12,030 | 7 | مختارات من كتب الأدب العربي | Arabic titles (multi-volume sets) |
| 4,781 | 16 | كتب عن تاريخ بلاد الحرمين و نجد - طبعات نادرة و قديمة و ملونة | Arabic titles |
| 3,350 | 10 | تاريخ التراث العربي - فؤاد سزكين | `tta####` codes |
A per-volume inventory of all 461 PDFs is published alongside this README as [`VOLUMES.md`](VOLUMES.md) (human-readable, grouped by collection) and [`volumes.json`](volumes.json) (machine-readable).
---
## All titled volumes (121)
Every PDF whose filename encoded a real book title. Pages sum to 45,013.
| pages | title |
|---:|---|
| 4,824 | خزانة الأدب ولب لباب لسان العرب |
| 2,547 | تجريد الأغاني |
| 2,123 | الكامل في اللغة والأدب |
| 2,013 | المقتصد في شرح التكملة للجرجاني |
| 1,676 | توضيح المقاصد والمسالك شرح ألفية ابن مالك للمرادي |
| 974 | كتاب الأمالي مع كتابي ذيل الأمالي والنوادر |
| 889 | أدب الكاتب |
| 792 | التقفية في اللغة |
| 696 | المنتخب من أدب العرب |
| 679 | الزاهر في غريب ألفاظ الشافعي |
| 669 | المعرب من كلام الاعجم على حروف المعجم |
| 654 | المدخل إلى تقويم اللسان |
| 596 | تصحيح الفصيح وشرحه |
| 583 | الفروق في اللغة |
| 582 | المقتضب |
| 557 | الدراسات اللغوية والنحوية في مؤلفات ابن تيمية |
| 552 | تاريخ أحمد باشا الجزار - حيدر أحمد الشهابي - مكتبة انطوان 1955 |
| 534 | الغريب المصنف |
| 510 | كتاب الاضداد |
| 503 | تاريخ نجد الحديث وملحقاته - دار ريحاني - بيروت 1954 |
| 498 | أبنية الأسماء والأفعال والمصادر، لابن القطاع الصقلي |
| 494 | أسرار العربية |
| 473 | اصلاح المنطق لابن السكيت |
| 471 | قلب جزيرة العرب - فؤاد حمزة - المطبعة السلفية ومكتبتها - مكة المكرمة - 1933 |
| 456 | جزيرة العرب في القرن العشرين ؛ لجنة التأليف والترجمة والنشر ؛ القاهرة 1354-1935 |
| 447 | تحفة المستفيد بتاريخ الاحساء في القديم والجديد 1 - آل عبد القادر - مطابع الرياض - الرياض - 1960 |
| 436 | شرح الفصيح |
| 425 | الموجز في قواعد اللغة |
| 420 | مختارات من أدب العرب |
| 418 | الدراسات النحوية واللغوية عن الزمخشري |
<details>
<summary><b>Show remaining 91 titled volumes</b></summary>
| pages | title |
|---:|---|
| 411 | لحن العوام |
| 398 | الاخطاء في حروف الجر |
| 393 | المشترك اللغوي |
| 385 | لغة قريش |
| 380 | تثقيف اللسان العربي |
| 378 | الفرق بين الاحرف الخمسة |
| 376 | التهذيب بمحكم الترتيب |
| 367 | البحث اللغوي عند العرب |
| 363 | ابن جني النحوي |
| 353 | النحو العصري |
| 342 | علم اللغة |
| 339 | صيحة في سبيل العربية |
| 339 | القراءة الراشدة |
| 338 | العشرات في اللغة |
| 324 | أسرار العربية (3) |
| 322 | أضواء على المملكة العربية السعودية - المديرية العامة للإذاعة - الرياض - 1958-1377 |
| 319 | ما اتفق لفظه واختلف معناه |
| 319 | الفروق اللغوية |
| 313 | معالم تاريخ الجزيرة العربية - سعيد عوض با وزير - مؤسسة الصبان - عدن 1966 |
| 313 | عمر فاروق الطباع - الوسيط في قواعد الإملاء والإنشاء |
| 305 | تثقيف اللسان وتلقيح الجنان |
| 299 | الجملة العربية |
| 280 | عنوان المجد في بيان أحوال بغداد والبصرة ونجد - دار منشورات البصري - بغداد 1962 |
| 279 | مناهج البحث في اللغة |
| 269 | الصاحبي في فقه اللغة |
| 268 | بحوث في اللغة |
| 268 | الشعر النبطي |
| 267 | التدريبات اللغوية |
| 265 | البلاد العربية السعودية - فؤاد حمزة, مطبعة أم القرى 1355-1937 |
| 263 | أخطاء اللغة عند المعاصرين |
| 262 | المملكة العربية السعودية في عهدها الحاضر - مديرية العامة للإذاعة, جدة 1376-1957 |
| 260 | اقتطاف الأزاهر |
| 254 | تقويم اللسان |
| 253 | تاريخ ملوك آل سعود - مطابع الرياض - 1961 |
| 239 | المطالع النصرية |
| 239 | تقويم اللسان (2) |
| 235 | أسرار العربية (2) |
| 227 | الصاحبي في فقه اللغة (2) |
| 224 | معجم الدخيل |
| 220 | الجملة العربية (2) |
| 219 | العربية والحداثة |
| 216 | المقتضب في لهجات العرب |
| 216 | الرياض المزهرة بين الكويت والمحمرة - عبد المسيح أنطاكي بيك - مطبعة العرب - القاهرة - 1907 |
| 213 | اللسان والإنسان |
| 208 | الحسن والإحسان |
| 199 | الشامل في اللغة لطلبة الجامعات والمعاهد غير المتخصصين |
| 197 | اللغة العربية أداءً ونطقاً |
| 177 | تعقبات مصابيح الجامع |
| 174 | المعرب من كلام الأعجمي على حروف المعجم |
| 172 | أسرار العربية (4) |
| 168 | أخطاء لغوية شائعة |
| 163 | ما تلحن فيه العامة |
| 162 | المقتبس من اللهجات |
| 153 | نحو تقويم جديد |
| 152 | عثرات اللسان |
| 146 | علم نفسك الخطوط |
| 146 | عنوان المجد في تاريخ نجد 1 ؛ مطبعة الشابندر 1328-1910 |
| 140 | الأمالي للزجاج |
| 139 | ملك وآمال الشعب - زينب الغزالي - دار الكتاب العربي؛ 1954 |
| 137 | الفرق بين الضاد والظاء للداني |
| 131 | زينة الفضلاء في الفرق بين الضاد والظاء |
| 130 | قواعد الإملاء |
| 127 | الكافي في اللغة |
| 127 | التشكيل الصوتي في اللغة |
| 110 | إتحاف الأماجد بما يصح به الاستشهاد |
| 108 | الفرق بين المذكر والمؤنث |
| 95 | شرح مثلثات قطرب |
| 94 | نقعة الصديان |
| 89 | التعريف بما آنست الهجرة من معالم دار الهجرة 1372-1952 |
| 77 | قواعد الإملاء (3) |
| 75 | كيف تكون فصيحاً |
| 67 | في قلب نجد والحجاز - مكتبة المنار - مصر 1927 |
| 64 | الإملاء والترقيم |
| 64 | الفرق بين الضاد والظاء |
| 64 | العربية |
| 53 | شرح الفصيح — مقدمة |
| 49 | الفرق بين الضاد والظاء (variant) |
| 46 | الزيادات على كتاب لحن العامة بالأندلس |
| 41 | كتاب السلاح |
| 28 | قواعد الإملاء (2) |
| 26 | التنبيه على حدوث التصحيف — مقدمة |
| 24 | أسرار العربية — مقدمة |
| 21 | الأرقام العربية |
| 20 | الفروق في اللغة — مقدمة |
| 11 | كتاب الأضداد — مقدمة |
| 10 | إصلاح المنطق لابن السكيت — مقدمة |
| 8 | إصلاح المنطق — نماذج |
| 8 | ما اتفق لفظه واختلف معناه — مقدمة |
| 5 | تثقيف اللسان العربي — مقدمة |
| 4 | مناهج البحث في اللغة — مقدمة |
| 1 | رسالتان في اللغة للأصمعي |
</details>
The remaining **339 PDFs** are filename-coded (not titled). Their codes are inventoried in [`VOLUMES.md`](VOLUMES.md) along with page counts and `pdf_url` links to their Archive.org source — you can cross-reference any individual code there to inspect it.
---
## Construction
1. **Merge.** 257 configs (`az_0000` through `az_0256`) of `HeshamHaroon/arabic-turath-ocr` — 500 rows each — concatenated into a single table. Pipeline bookkeeping columns (`clean_model`, `cleaned`, `error`) dropped.
2. **Normalize text.** `text_norm` derived from `text` via:
- Unicode NFC normalization
- removal of U+0640 (ـ, tatweel)
- removal of U+064B–U+065F and U+0670 (tashkeel / harakat)
- expansion of lam-alef ligatures (ﻻ → لا, ﻷ → لأ, ﻹ → لإ, ﻵ → لآ)
- collapse of all whitespace — including zero-width joiner, zero-width non-joiner, and NBSP — to a single space
3. **Filter.** Rows were dropped where:
- `char_count_norm < 50` (OCR fragments, blank pages)
- `arabic_ratio_norm < 0.5` (foreign-language pages, tables, colophons)
- `text_norm` exactly duplicates an earlier row (ordered by `book, page`)
4. **Fix column semantics.** In the upstream dataset, `item_title` actually held the Archive.org collection name and `book` held the PDF filename stem (sometimes prefixed with volume/part/issue markers). Those columns were renamed to `collection` and `book_raw`, and a derived `book_title` (+ `book_title_source` provenance flag) was added. Filenames consisting purely of letters-then-digits (e.g. `a025`, `tta0104`) are classified as `pdf_code`.
5. **Yield.** 123,062 of 128,105 input rows kept — 96.1%.
### Drops breakdown
| reason | rows |
|---|---:|
| empty / < 50 chars after normalization | 3,730 |
| Arabic-letter ratio < 0.5 | 1,128 |
| exact-duplicate normalized text | 185 |
### Arabic-letter ratio distribution (of merged input)
| bucket | rows |
|---|---:|
| < 30% Arabic | 3,208 |
| 30–50% | 415 |
| 50–70% | 1,085 |
| 70–90% | 7,933 |
| ≥ 90% Arabic | 115,464 |
### book_title_source breakdown
| source | rows | PDFs |
|---|---:|---:|
| `parsed_from_pdf_name` | 45,013 | 122 |
| `pdf_code` | 78,049 | 339 |
---
## Suggested use
- **Causal LM pre-training / continued pre-training** on classical Arabic — `text_norm`.
- **OCR denoising**: `raw_text``text` as noisy-clean training pairs.
- **Lexical and grammatical resources**: the top ten volumes are dictionaries, fiqh al-lugha works, and commentaries.
- **Domain adapters** for models that need to handle pre-modern Arabic registers.
For tasks where diacritics or glyph form matter — Quranic recitation alignment, classical prosody, tashkeel restoration — use `text` or `raw_text` rather than `text_norm`.
---
## Licensing & provenance
All underlying PDFs are public-domain scans of classical Arabic works, hosted on the Internet Archive. The aggregation, cleaning, and metadata work in this repository are released under **CC-BY-4.0**. When quoting or reproducing content, cite the original book (see `book_title`) and, where helpful, the Archive.org item via `source_detail_url`.
---
## Citation
```bibtex
@dataset{haroon_asfar_2026,
author = {Haroon, Hesham},
title = {Asfar: An Arabic Heritage Text Corpus},
year = {2026},
url = {https://huggingface.co/datasets/HeshamHaroon/asfar},
}
```