orature/ALIF_Urdu_Corpus_AUC

ALIF_Urdu_Corpus (Preview)

This dataset, ALIF_Urdu_Corpus, is part of the ALIF الف project by Orature AI. It was curated for pretraining Urdu language models. It serves as a preview to our entire 33GB Dataset.

Purpose of the Dataset:

• (preview) To serve as a large-scale, diverse, and high-quality foundation for pretraining generative language models for Urdu.

Languages

The data is in Urdu.

Dataset Structure

(For Pretraining Corpus - ALIF-Urdu-Corpus): The dataset is structured as a collection of text entries, in CSV format, with the following columns:

Data	Category	Source

• Data: The Data column contains the actual Urdu data
• Category: The Category column refers to the type of data it is, for example CommonCrawl, Fineweb, etc.
• Source: The Source column contains the actual source from where the data was taken.

Data Collection and Preprocessing

The Complete ALIF-Urdu-Corpus: The dataset was meticulously collected from a variety of sources to ensure diversity and coverage:

• Common Crawl Dumps: 11.3 GB (Dump 1) and 8.1 GB (Dump 2) of filtered Urdu text.
• Translation Data: 5.5 GB of educational content from the English FineWeb dataset, translated to Urdu using Google Translate API.
• News Websites: 3.3 GB scraped from various Urdu news websites.
• Existing Datasets: 2.9 GB from public Urdu corpora (e.g., UrduHack, other open-source).
• Books (OCR Processed): 1.3 GB of text extracted from scanned Urdu books using Google Vision OCR, followed by post-OCR cleaning.
• Blog Sites: 0.6 GB from various Urdu blogs.

Preprocessing Steps:

1. Cleaning: Removal of HTML tags, links, numbers (unless contextually relevant), email addresses, and other non-linguistic noise.
2. Encoding Normalization: Ensured consistent UTF-8 encoding.
3. Language Filtering: Non-Urdu content was filtered out using language detection tools.
4. Deduplication: Rigorous deduplication was performed using MinHash-based Locality Sensitive Hashing (LSH) to identify and remove near-duplicate documents and paragraphs, both within and across source datasets. Exact duplicates were also removed.
5. Formatting: Final data organized into a structured format (e.g., CSV), with End-of-Text (EOT) tokens used to delineate documents/segments during training.

Dataset Size

• ALIF_Urdu_Corpus:
  • Total Size: ~33 GB for ALIF-Urdu-Corpus, however, this dataset preview contains about 13.7MB of that data.
  • Number of Rows/Examples: 5000 rows

Intended Uses

• Pretraining Language Models: The ALIF-Urdu-Corpus is primarily intended for pretraining large-scale generative language models for Urdu.
• Instruction Fine-tuning: The ALIF-Urdu-Instruct dataset is designed for fine-tuning pretrained models to follow instructions in Urdu.
• NLP Research: Can be used for various research tasks in Urdu NLP, such as studying linguistic phenomena, bias in text, or developing new preprocessing techniques.
• Benchmarking: Subsets can be used for creating benchmarks for Urdu language understanding or generation.