3arabLM-corpus

A large-scale tokenized corpus of Arabic Islamic texts from the Shamela Library, pre-tokenized and ready for training Arabic language models and building RAG (Retrieval-Augmented Generation) systems. This is the first batch—comprising about 521 books out of a total of 35,000 and I will be adding them successively.

📊 Statistics

Metric	Value
Number of files	47 Parquet files
Total rows (chunks)	2,209,875
Sequence length	1,024 tokens per chunk
Total tokens	≈ 2,262,912,000 (2.26B)
Total size	~3.07 GB
Unique books	521+
Categories	8 main categories
Language	Modern Standard Arabic (classical Islamic texts)

---

🗂️ Data Structure

Each Parquet file contains only two columns (since texts are pre-tokenized):

Column	Type	Description
input_ids	list<int32>	1,024 token IDs per row
attention_mask	list<int8>	1 for real tokens, 0 for padding

Tokenizer

• Path: finsh/v2 (from model sherif1313/3arabLM-4B-islamic-v2)
• Vocabulary size: 248,044
• EOS token ID: 248,044
• PAD token ID: 248,044

---

📚 Categories

Category	Approximate Share
Tafsir (Quranic Exegesis)	56%
Hadith Texts	22%
Fatwas (Islamic Rulings)	7%
General Fiqh (Jurisprudence)	5%
Ibn Taymiyyah's Works	4%
Arabic Grammar & Morphology	2%
Aqeedah (Creed)	2%
Ibn al-Qayyim's Works	1%

---

📖 Sample Books All books are defined in a file Book.txt

General Fiqh

• Al-Mufassal fi Fiqh al-Jihad
• Al-Fiqh al-Islami wa Adillatuh (Al-Zuhayli)
• Al-Muhalla bil-Athar
• Kuwaiti Fiqh Encyclopedia
• Fiqh al-Sunnah
• Fiqh ala al-Madhahib al-Arba'a

Fatwas

• Fatwas of the Permanent Committee - Vol 1
• Fatwas of the Permanent Committee - Vol 2
• Nur ala al-Darb Fatwas (Ibn Baz)
• Majmu' Fatawa wa Rasa'il al-Uthaymeen
• Majmu' Fatawa Ibn Baz
• Fatwas of the Islamic Network
• Islam Question and Answer Website

Ibn Taymiyyah's Works

• Majmu' al-Fatawa
• Al-Jawab al-Sahih
• Al-Furqan bayna Awliya' al-Rahman wa Awliya' al-Shaytan
• Al-Risalah al-Arshiyyah
• Al-Safadiyyah
• Al-Mustadrak ala Majmu' al-Fatawa

Ibn al-Qayyim's Works

• I'lam al-Muwaqqi'in
• Bada'i' al-Fawa'id
• Madarij al-Salikin
• Zad al-Ma'ad
• Al-Tibb al-Nabawi
• Al-Wabil al-Sayyib
• Al-Da' wa al-Dawa'

Tafsir (Quranic Exegesis)

• Tafsir al-Tabari
• Tafsir al-Qurtubi
• Tafsir Ibn Kathir
• Tafsir al-Razi (Mafatih al-Ghayb)
• Al-Bahr al-Madid
• Al-Durr al-Manthur
• Fi Zilal al-Quran
• Al-Tahrir wa al-Tanwir

Hadith Texts

• Sahih al-Bukhari
• Sahih Muslim
• Sunan al-Tirmidhi
• Sunan al-Darimi
• Musnad Ahmad
• Al-Sunan al-Kubra (Al-Bayhaqi)
• Al-Mu'jam al-Kabir (Al-Tabarani)

Arabic Grammar & Morphology

• Al-Nahw al-Wafi
• Sharh al-Ashmouni
• Awdah al-Masalik
• Sharh al-Mufassal (Ibn Ya'ish)
• Al-Khasa'is
• Ham' al-Hawami'

Aqeedah (Creed)

• Hayat al-Khidr wa Mawtuh
• Al-Jami' li Shu'ab al-Iman
• Risalah fi Haqiqat al-Ta'wil
• Al-Iman bil-Qadar
• Al-Mawaqif (Al-Iji)

---

📁 File-to-Category Mapping

Each Parquet file corresponds to one or more books. Since books are packed sequentially, files marked with ↳ are continuations of the previous file's book (no new book header at the start).

File	Category
packed_ids_0000.parquet	فقه عام (General Fiqh)
packed_ids_0001.parquet	فقه عام (General Fiqh)
packed_ids_0002.parquet	فقه عام (General Fiqh)
packed_ids_0003.parquet	الفتاوى (Fatwas)
packed_ids_0004.parquet	الفتاوى (Fatwas)
packed_ids_0005.parquet	↳ الفتاوى (continuation)
packed_ids_0006.parquet	كتب ابن القيم (Ibn al-Qayyim)
packed_ids_0007.parquet	كتب ابن تيمية (Ibn Taymiyyah)
packed_ids_0008.parquet	كتب ابن تيمية (Ibn Taymiyyah)
packed_ids_0009.parquet	النحو والصرف (Grammar & Morphology)
packed_ids_0010.parquet	↳ النحو والصرف (continuation)
packed_ids_0011.parquet	التفاسير (Tafsir)
packed_ids_0012.parquet	↳ التفاسير (continuation)
packed_ids_0013.parquet	↳ التفاسير (continuation)
packed_ids_0014.parquet	↳ التفاسير (continuation)
packed_ids_0015.parquet	↳ التفاسير (continuation)
packed_ids_0016.parquet	↳ التفاسير (continuation)
packed_ids_0017.parquet	↳ التفاسير (continuation)
packed_ids_0018.parquet	↳ التفاسير (continuation)
packed_ids_0019.parquet	↳ التفاسير (continuation)
packed_ids_0020.parquet	↳ التفاسير (continuation)
packed_ids_0021.parquet	↳ التفاسير (continuation)
packed_ids_0022.parquet	↳ التفاسير (continuation)
packed_ids_0023.parquet	↳ التفاسير (continuation)
packed_ids_0024.parquet	↳ التفاسير (continuation)
packed_ids_0025.parquet	↳ التفاسير (continuation)
packed_ids_0026.parquet	التفاسير (Tafsir)
packed_ids_0027.parquet	التفاسير (Tafsir)
packed_ids_0028.parquet	التفاسير (Tafsir)
packed_ids_0029.parquet	↳ التفاسير (continuation)
packed_ids_0030.parquet	↳ التفاسير (continuation)
packed_ids_0031.parquet	التفاسير (Tafsir)
packed_ids_0032.parquet	التفاسير (Tafsir)
packed_ids_0033.parquet	↳ التفاسير (continuation)
packed_ids_0034.parquet	↳ التفاسير (continuation)
packed_ids_0035.parquet	↳ التفاسير (continuation)
packed_ids_0036.parquet	متون الحديث (Hadith Texts)
packed_ids_0037.parquet	↳ متون الحديث (continuation)
packed_ids_0038.parquet	↳ متون الحديث (continuation)
packed_ids_0039.parquet	متون الحديث (Hadith Texts)
packed_ids_0040.parquet	↳ متون الحديث (continuation)
packed_ids_0041.parquet	متون الحديث (Hadith Texts)
packed_ids_0042.parquet	متون الحديث (Hadith Texts)
packed_ids_0043.parquet	↳ متون الحديث (continuation)
packed_ids_0044.parquet	متون الحديث (Hadith Texts)
packed_ids_0045.parquet	↳ متون الحديث (continuation)
packed_ids_0046.parquet	العقيدة (Aqeedah / Creed)

Summary by Category

Category	Files	File Range
فقه عام (General Fiqh)	3	0000–0002
الفتاوى (Fatwas)	3	0003–0005
كتب ابن القيم (Ibn al-Qayyim)	1	0006
كتب ابن تيمية (Ibn Taymiyyah)	2	0007–0008
النحو والصرف (Grammar)	2	0009–0010
التفاسير (Tafsir)	22	0011–0035
متون الحديث (Hadith)	10	0036–0045
العقيدة (Aqeedah)	1	0046

⚠️ Important Notes

1. Files marked with ↳ do not start with a new book header. They are continuations of the previous file's book. A single book can span multiple Parquet files when its token count exceeds 50,000 chunks.

2. The category is NOT stored inside the Parquet file. Each file has only input_ids and attention_mask columns. To determine a chunk's category, you must decode the text and look for the header: الكتاب: المؤلف: التصنيف:

text

3. Each file has ~50,000 rows (except the last file of each book, which may be smaller).

4. To extract a specific category, use the file range from the table above.

Example: Extract All Tafsir Chunks

import pandas as pd

# Tafsir is in files 0011 to 0035
tafsir_files = [f"data/packed_ids_{i:04d}.parquet" for i in range(11, 36)]
tafsir_chunks = pd.concat([pd.read_parquet(f) for f in tafsir_files])

print(f"Total Tafsir chunks: {len(tafsir_chunks):,}")
Example: Extract All Hadith Chunks
python
# Hadith is in files 0036 to 0045
hadith_files = [f"data/packed_ids_{i:04d}.parquet" for i in range(36, 46)]
hadith_chunks = pd.concat([pd.read_parquet(f) for f in hadith_files])

print(f"Total Hadith chunks: {len(hadith_chunks):,}")
Example: Generate a Full File-to-Category Map
python
import glob, json, re
import pyarrow.parquet as pq
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("sherif1313/3arabLM-4B-islamic-v2")

HEADER_RE = re.compile(
    r"الكتاب:\s*([^\n]+)\n"
    r"المؤلف:\s*([^\n]+)\n"
    r"التصنيف:\s*([^\n]+)"
)

mapping = {}
for f in sorted(glob.glob("data/packed_ids_*.parquet")):
    pf = pq.ParquetFile(f)
    tbl = pf.read_row_group(0, columns=["input_ids", "attention_mask"])
    ids  = tbl.column("input_ids")[0].as_py()
    mask = tbl.column("attention_mask")[0].as_py()
    text = tok.decode([i for i, m in zip(ids, mask) if m], skip_special_tokens=True)

    m = HEADER_RE.search(text[:1500])
    mapping[f.split("/")[-1]] = {
        "book":     m.group(1).strip() if m else None,
        "author":   m.group(2).strip() if m else None,
        "category": m.group(3).strip() if m else None,
    }

with open("file_categories.json", "w", encoding="utf-8") as fp:
    json.dump(mapping, fp, ensure_ascii=False, indent=2)

🚀 Usage Load the dataset

from datasets import load_dataset

ds = load_dataset("sherif1313/3arabLM-corpus", split="train")
print(ds)
# Dataset({
#     features: ['input_ids', 'attention_mask'],
#     num_rows: 2209875
# })
Decode a sample
python
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained(
    "sherif1313/3arabLM-4B-islamic-v2",
    trust_remote_code=True,
)

sample = ds[0]
ids  = sample["input_ids"]
mask = sample["attention_mask"]


# Extract only real tokens (no padding)
real_ids = [i for i, m in zip(ids, mask) if m]
text = tok.decode(real_ids, skip_special_tokens=True)
print(text[:500])
Read a single file directly

import pandas as pd

df = pd.read_parquet("data/packed_ids_0000.parquet")
print(df.head())
#    input_ids                                        attention_mask
# 0  [30350, 214854, 25, 174721, ...]                [1, 1, 1, ...]

🏗️ File Structure Each Parquet file contains 50,000 rows (except the last one), where each row = 1,024 tokens.

During packing, the following header was prepended to each chunk:

text Book: Author: Category: Part:

[PAGE ]

🎯 Use Cases Training Arabic language models (Pre-training / Continued Pre-training)

Building Arabic RAG systems using FAISS + Arabic embedding models

Semantic search in classical Islamic and Arabic texts

Analyzing Islamic texts (Fiqh, Hadith, Tafsir, Aqeedah)

Metadata-Guided Generation (source-grounded text generation)

🔗 Related Repositories

Repository Description sherif1313/3arabLM-4B-islamic-v2 3arabLM 4B model trained on Shamela sherif1313/3arabLM-RAG Pre-built FAISS index (coming soon) 📄

The original texts are sourced from the Shamela Library Official Library: goldenshamela

Tokenization and structure: Apache 2.0