Datasetpublic

Minuri/sinhala-corpus-a-news-1m

News-Only Sinhala Corpus A news-domain subset of 1M Sinhala sentences sampled from the Minuri/diverse_sinhala_dataset corpus, used for continual pretraining of LLaMA 3.2 1B (Model A) as part of a diversity-driven Sinhala language model adaptation study at the Informatics Institute of Technology (IIT), Colombo, affiliated with Robert Gordon University (RGU). Corpus variants in this series: Minuri/sinhala-corpus-a-news-1m - News-only subset (domain-homogeneous baseline) - this… See the full description on the dataset page: https://huggingface.co/datasets/Minuri/sinhala-corpus-a-news-1m.

sourceHugging Facecc-by-sa-4.0updated 6mo agoView on Hugging Face

0likes27downloads

Dataset Card

News-Only Sinhala Corpus

A news-domain subset of 1M Sinhala sentences sampled from the Minuri/diverse_sinhala_dataset corpus, used for continual pretraining of LLaMA 3.2 1B (Model A) as part of a diversity-driven Sinhala language model adaptation study at the Informatics Institute of Technology (IIT), Colombo, affiliated with Robert Gordon University (RGU).

Corpus variants in this series: - Minuri/sinhala-corpus-a-news-1m - News-only subset (domain-homogeneous baseline) - this repo - Minuri/sinhala-corpus-b-random-1m - Random subset (random baseline) - Minuri/sinhala-corpus-c-diverse-1m - Diversity-optimized subset ✅ Best perplexity

Dataset Description

Corpus A serves as the domain-homogeneous baseline, comprising sentences drawn exclusively from the news domain of the parent corpus. This enables controlled comparison against the random (B) and diversity-optimized (C) corpora in downstream perplexity and evaluation experiments. The model trained on this corpus (Model A) achieved a perplexity of 14.68 on the Sinhala test set.

Source Datasets (via parent corpus)

Source	Description
`culturax`	CulturaX multilingual web corpus (Sinhala subset)
`nsina`	NSina Sinhala news corpus
`madlad`	MADLAD-400 multilingual dataset (Sinhala subset)
`wikipedia`	Sinhala Wikipedia dump

Dataset Structure

Column	Type	Description
`orig_index`	int	Original index in the parent corpus
`sentence`	string	Sinhala sentence text
`source`	string	Source dataset identifier
`predicted_domain`	string	Domain label predicted by XLM-RoBERTa classifier
`confidence`	float	Classifier confidence score

Splits

Split	Rows
train	1,000,000

Format

Available in both JSONL and CSV formats.

Intended Uses

—Continual pretraining of LLMs on Sinhala (domain-homogeneous baseline)
—Ablation studies on corpus diversity
—Sinhala NLP benchmarking

Associated Model

This corpus was used to train: Minuri/sinhala-llama-1b-corpus-news

Sources & Licenses

This dataset contains sentences derived from the following source datasets. Users must comply with the license terms of each:

Source	License	Notes
MADLAD-400	ODC-BY	Attribution required
CulturaX	mC4 + OSCAR licenses	Requires contact info agreement on HuggingFace before access
wikimedia/wikipedia	CC BY-SA 3.0 + GFDL	ShareAlike - derived works must carry same license
sinhala-nlp/NSINA	CC BY-SA 4.0	ShareAlike - derived works must carry same license

This dataset is released under CC BY-SA 4.0 in compliance with the ShareAlike terms of Wikipedia and NSINA.