wafaali786/sindhi-text-corpus

Sindhi Cleaned Text Corpus

A cleaned, de-duplicated corpus of Sindhi (Perso-Arabic script) text for language-model pretraining and research on a low-resource language.

Contents

File	Lines	Words
train.txt	1,016,307	48,245,742
validation.txt	10,000	479,094
test.txt	10,000	475,594
Total (410.3 MB)	1,036,307	49,200,430

Plain UTF-8 text, one cleaned line per row. No labels. The split is a random line-level shuffle with seed 42.

Sources

• CC-100 Sindhi (https://data.statmt.org/cc-100/): web-crawled text derived from Common Crawl. 967,904 lines kept after cleaning.
• Sindhi Wikipedia (wikimedia/wikipedia, config 20231101.sd): 68,403 lines kept after cleaning.

Cleaning steps

1. Unicode normalization (NFKC)
2. Removal of URLs and email addresses
3. Removal of invisible characters and tatweel
4. Removal of a trailing English site name (for example "- Daily Hilal-e-Pakistan")
5. Dropping lines shorter than 4 words
6. Dropping lines where fewer than 80% of letters are Arabic script
7. Removing exact duplicate lines

Intended use

Pretraining or continued pretraining of language models, tokenizer training, and language-modelling research on Sindhi. Not suited for tasks that need labels without adding annotations.

Example result

XLM-RoBERTa base, trained for 3,000 steps on part of this corpus, reached a test perplexity of 7.37 (original model: 11.83). Its top-1 accuracy at guessing a hidden word went from 47.0% to 58.9% on 1,000 test questions.

Known limitations

• The split is random by line, so lines from the same article can appear in both train and test.
• Some site names remain at the end of lines (for example "_ KTN NEWS").
• The script filter does not separate Sindhi from other languages that use the same script (for example Urdu), and no language-identification model was used.
• Mostly news, web and Wikipedia text; literature, spoken language and social media are barely covered.
• Not screened for personal or offensive content.

License and attribution

This is a mixed-source corpus, so no single standard license applies.

• Sindhi Wikipedia text is available under CC BY-SA 4.0.
• CC-100 is derived from Common Crawl web pages. Copyright in the original pages stays with the website owners. Use of CC-100 follows the Common Crawl terms of use.

Provided for research and educational use. Please check each source's current terms before redistributing, and cite the sources:

• CC-100: Conneau et al. (2020), "Unsupervised Cross-lingual Representation Learning at Scale"; Wenzek et al. (2020), "CCNet: Extracting High Quality Monolingual Datasets from Web Crawl Data".
• Wikipedia: https://dumps.wikimedia.org/