sherrycodes911/Sindhi-NLP-Corpus

Sindhi NLP Corpus

A production-grade Sindhi NLP corpus engineered by Shahriyar.

Dataset Details

A dataset for the Sindhi language (snd). This corpus is engineered for Large Language Model (LLM) pre-training, fine-tuning, NLP benchmarking, and any broad computational or linguistic purpose involving Sindhi text, addressing the severe lack of high-quality, clean data for low-resource languages.

• Curated by: Shahriyar
• Language(s) (NLP): Sindhi
• License: CC-BY-SA-4.0

Uses

This dataset is exported into specific formats to support different workflows in the community:

If you are:

• Data Scientists, grab the .parquet to run fast analytics and statistical modeling.
• LLM Engineers, grab the .jsonl to pass directly into fine-tuning APIs (like OpenAI, Anthropic, or local instruction-tuning scripts).
• NLP Researchers, use the .arrow (via Hugging Face datasets) to stream the data directly into your PyTorch, TensorFlow, or JAX training loops.

Out-of-Scope Use

This dataset is a generalized text corpus. It should not be used as a definitive source of factual ground truth without verification. Because it contains web-sourced and platform-sourced text, it is not guaranteed to be free from societal biases, and should not be used to train systems intended to generate harmful, toxic, or discriminatory content.

Dataset Structure

The corpus consists of two columns:

• text: Contains the cleaned, exact-deduplicated Sindhi text sequence.
• source: A uniform tracking label (sindhi_corpus) applied to all rows

Dataset Statistics

• Total Cleaned Sequences: 5,743,268
• Total Words: 344,674,535
• Estimated Subword Tokens: ~600M - 800M (Depending on BPE/SentencePiece tokenizer)
• Average Sentence Length: 6.04 words

Dataset Creation

Curation Rationale

The primary motivation for this dataset is the severe under-representation of the Sindhi language in modern NLP and Generative AI. Existing open-source datasets are notoriously noisy, heavily duplicated, and fragmented. This benchmark unifies these fragmented sources into a single, high-signal, production-ready corpus.

Source Data

Data Collection and Processing

To build this benchmark, an extensive raw corpus totaling over 57 million rows was meticulously collected. This raw data was sourced from a wide variety of authentic Sindhi digital platforms and open-source datasets.

The data was processed through a strict pipeline to remove noise, correct schemas, normalize encodings, and perform aggressive exact-deduplication using cryptographic hashing. The final result successfully boiled the 57 million raw rows down to 5.7 million highly unique, high-signal sequences.

Bias, Risks, and Limitations

While rigorous cleaning and exact-deduplication were applied, the data originates from public internet sources. As such, it naturally carries the dialectical patterns, societal biases, and opinions present in the original texts. Users training large language models should be aware that the data may contain implicit biases or historical perspectives tied to the source materials.

Quickstart

You don't need to download the files locally to use them. You can stream the dataset directly using the Hugging Face datasets library:

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("sherrycodes911/Sindhi-NLP-Corpus")

# Sample the first sequence
print(dataset['train'][0]['text'])