aakashMeghwar01/sindhi-corpus-505m

Sindhi Corpus 505M

The largest open-source, deduplicated Sindhi language pretraining corpus.

~505 million tokens across 742K documents, covering news, literature, legal, religious, encyclopedic, and web-crawled Sindhi text. Built for training Sindhi language models, tokenizers, and NLP tools.

Dataset Summary

Stat	Value
Documents	~742,000
Tokens (estimated)	~505 million
Language	Sindhi (sd) — Arabic script
Format	Parquet (single text column)
License	Apache 2.0
Curator	Aakash Meghwar

How to Use

from datasets import load_dataset

# Full dataset
ds = load_dataset("aakashMeghwar01/sindhi-corpus-505m", split="train")
print(ds[0]["text"][:200])

# Streaming (recommended for Colab/Kaggle)
ds = load_dataset("aakashMeghwar01/sindhi-corpus-505m", split="train", streaming=True)
for example in ds:
    print(example["text"][:200])
    break

Source Datasets

This corpus was compiled from 11 publicly available Sindhi datasets spanning diverse genres:

#	Source	Type	URL
1	AMBILE Sindhi Mega Corpus	Mixed (news, web, literature)	Kaggle
2	CC100-Sindhi	Web crawl (Common Crawl)	Metatext
3	Daily Kawish Articles	Newspaper articles	Kaggle
4	Sindhi News — Awami Awaz	News articles	Kaggle
5	Sindhi News — Sindh Express	News articles	Kaggle
6	Encyclopedia Sindhiana	Encyclopedic articles	Kaggle
7	Sindhi Legal Dataset	Legal documents	Kaggle
8	Sindhi Religious Data	Religious texts	Kaggle
9	Sindhi Language Corpus	Mixed	Kaggle
10	Sindhi Stopwords	Linguistic resource	Kaggle
11	Zenodo Sindhi Dataset	Mixed	Zenodo

Data Collection & Processing Pipeline

The corpus was assembled through a multi-stage pipeline:

Stage 1: Collection

Raw text was extracted from all 11 sources. CSV files were parsed with automatic column detection (looking for text, content, article columns). TXT files were split on paragraph boundaries. All text was read as UTF-8.

Stage 2: Normalization

• Unicode NFC normalization to canonicalize Arabic-script characters
• Encoding repair via ftfy (fixes mojibake and broken encodings)
• Character variant unification: Arabic Yeh → Farsi Yeh, Arabic Kaf → Keheh, Arabic Heh → Heh Goal (standard Sindhi forms)
• URL and email removal
• HTML tag stripping
• Whitespace normalization

Stage 3: Quality Filtering

Documents were filtered based on:

• Minimum length: 50 characters (removes fragments)
• Maximum length: 100,000 characters (removes data dumps)
• Sindhi script ratio: At least 30% of non-space characters must be Arabic-script (removes code-mixed noise, English-only documents)
• Repetition check: Documents with <10% unique characters rejected (catches OCR errors, repetitive noise)

Stage 4: Deduplication

Two-stage deduplication to remove near-identical documents:

1. Exact dedup: MD5 hash of full document text
2. Near-duplicate detection: MinHash LSH with 128 hash functions, 5-word shingles, Jaccard threshold 0.85

Stage 5: Final Assembly

• Documents shuffled randomly
• Exported as Parquet with a single text column
• Pushed to HuggingFace Hub

Processing Tools

• pandas for CSV/data handling
• ftfy for encoding repair
• datasketch (MinHash LSH) for near-duplicate detection
• unicodedata for Unicode normalization
• Custom SindhiTextProcessor class with Sindhi-specific regex patterns

Dataset Composition

The corpus covers diverse domains to ensure broad linguistic coverage:

Domain	Sources	Coverage
News	Daily Kawish, Awami Awaz, Sindh Express	Current affairs, politics, sports, editorials
Encyclopedia	Encyclopedia Sindhiana	History, culture, geography, biography
Web Crawl	CC100-Sindhi	Diverse web text, blogs, forums
Legal	Sindhi Legal Dataset	Legal documents, court proceedings
Religious	Sindhi Religious Data	Religious texts and commentary
Mixed	AMBILE Mega Corpus, Sindhi Language Corpus, Zenodo	Literature, academic, general

Intended Use

This corpus is designed for:

• Pretraining Sindhi language models (GPT, BERT, etc.)
• Tokenizer training (BPE, Unigram, SentencePiece)
• NLP tool development (stemmer evaluation, stopword extraction, morphological analysis)
• Linguistic research on Sindhi text

Models Trained on This Corpus

Model	Type	Params	Link
SindhiLM	GPT-2 from scratch	37.8M	HF
SindhiLM-Qwen-0.5B	Qwen2.5-0.5B fine-tune	0.5B	HF
SindhiLM-Qwen-0.5B-v2	Qwen2.5-0.5B SFT	0.5B	HF
SindhiLM-Tokenizer-v2	Morpheme-aware BPE	—	HF

Associated Tools

Tool	Description	Link
SindhiNLTK	Morphology-aware NLP toolkit for Sindhi	PyPI · GitHub

Tokenized Version

A pre-tokenized version of this corpus is also available:

• sindhi-tokenized-505m — Same corpus, tokenized with SindhiLM-Tokenizer-v1

Limitations & Biases

• News-heavy: Newspaper sources (Kawish, Awami Awaz, Sindh Express) form a significant portion, which may bias toward formal journalistic Sindhi.
• Script variants: Despite normalization, some character variant inconsistencies may remain in web-crawled text.
• Temporal bias: News articles are concentrated around their publication dates; no uniform temporal sampling was applied.
• Deduplication residual: Near-duplicate detection at threshold 0.85 may leave some paraphrased duplicates.

Citation

@dataset{meghwar2026sindhi505m,
  author    = {Aakash Meghwar},
  title     = {Sindhi Corpus 505M: A Deduplicated Pretraining Dataset for Sindhi},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/aakashMeghwar01/sindhi-corpus-505m}
}

Contact

Aakash Meghwar — Computational Linguist

• HuggingFace: aakashMeghwar01
• GitHub: AakashKumarMissrani