Kashif786/sindhi-gold-corpus-mlm

📚 Sindhi Gold Corpus (MLM Optimized) prepared from (arnizamani/Sindhi-texts-big-dataset)

The Sindhi Gold Corpus is a highly curated, mathematically balanced, and rigorously sanitized dataset containing 1.67 GB of pure Sindhi text. Comprising exactly 5,424,753 continuous lines, this corpus was specifically engineered for continued pre-training and Masked Language Modeling (MLM) of transformer architectures (such as mBERT, ELECTRA, and DistilBERT).This dataset is a surgically refined derivative of the massive parent dataset, arnizamani/Sindhi-texts-big-dataset (Version 2.0). The parent corpus is a monumental 3.19 GB collection containing approximately 1.81 billion characters and 390,000 documents across 9 distinct sources. Assembled by Abdul Rahim Nizamani to support low-resource language research, the original dataset combines material digitized by Sindhi literary institutions and forums, a Sindhi encyclopedia, newspaper archives, a classical dictionary, and the Sindhi portions of two large web-crawl corpora (FineWeb2 and CC100).
Instead of indiscriminately merging all 3.19 GB of available text from the parent dataset, this "Gold Standard" distribution heavily weights the human-curated literature and journalism from the original sources, while strictly capping the noisy web crawls to prevent model degradation.

📊 Dataset Statistics

• Total File Size: 1676.22 MB (~1.67 GB)
• Total Valid Lines: 5,424,753
• Format: Hugging Face text / Parquet (1 feature: text)
• Language: Sindhi (snd_Arab)

🎯 The "Gold Standard" Curation Strategy

In low-resource NLP, dumping raw internet crawls into a model often destroys contextual understanding. To ensure the model learns deep grammatical structures rather than web boilerplate, this corpus employs a strict volume and quality distribution strategy:

1. 100% Curated Literature & Journalism: Retained every single line of high-quality, human-written text (Books, Adabi Board, Dictionaries, and Newspapers).
2. 20% Web Crawl Cap: Randomly downsampled the massive FineWeb2 and CC100 datasets to exactly 20%. This injects just enough modern internet context and contemporary vocabulary without letting algorithmic noise overwhelm the clean data.
3. Translation Balancing: Downsampled the 14 duplicate Quranic translations to just 2, preventing the embedding space from becoming overly biased toward specific religious phrasing.

Corpus Breakdown by Source

Source	Retained Ratio	Valid Lines	Content Profile
Sindh Salamat Books	100%	1,999,584	Deep paragraph structures (Avg 30 words/line)
Sindhi Adabi Board	100%	1,953,509	Grammatical backbone (Avg 46 words/line)
FineWeb2 (Web Crawl)	20%	965,314	Contemporary internet text and slang
CC100-Sindhi (Web Crawl)	20%	271,963	General web context
Jamesindhilughaat	100%	81,661	Dictionary definitions (Avg 18 words/line)
Sindhiana.org	100%	81,013	Factual encyclopedia data
Newspapers	100%	59,237	Dense journalistic prose (Avg 184 words/line)
Quran Translations	2 of 14	12,472	Classical structural text
Total		5,424,753

🧹 Data Purification & Hygiene

Raw datasets often contain artifacts that pollute a model's embedding matrix. This corpus underwent a strict sanitization pipeline prior to upload:

• UI Boilerplate Eradication: Swept and removed 19,230 web-scraping artifacts that were stamped across book files. The model will not waste capacity learning website menus like "Index", "Text size", "Change font", or "سنڌسلامت رِيڊ 2.0".
• Numerical Indexing Removal: Surgically stripped 12,472 instances of rigid numerical indexing (e.g., 1|1|) from religious texts using Regex, converting them into pure natural language strings to prevent the model from hallucinating chapter/verse numbers.
• Empty Line Deletion: Removed all blank lines to ensure unbroken 512-token contiguous chunking during the MLM grouping phase.

🚀 Intended Use Case: Masked Language Modeling

This dataset is structurally flawless for MLM. Because the newspaper and book segments average between 30 and 184 words per line, the text is inherently rich in context. It is designed to be easily chunked into dense BLOCK_SIZE = 512 sequences without massive padding inefficiencies.

How to load directly into a Hugging Face Trainer:

from datasets import load_dataset

# Loads the entire 1.67 GB corpus natively
dataset = load_dataset("Kashif786/sindhi-gold-corpus-mlm", split="train")