Open datasets

Open, citable, and reproducible.

All lab datasets are officially published on Harvard Dataverse with permanent DOIs, alongside companion models on HuggingFace. Each release includes detailed data cards, standardized loading scripts, and fully documented annotation methodology.

Roman Urdu Sentiment Corpus (RUDaSA)

Harvard Dataverse · HuggingFace

Sentiment AnalysisRoman UrduLow-Resource NLPHarvard Dataverse

A large-scale Roman Urdu sentiment dataset built via privacy-preserving embedding pipelines. Competitive benchmarking of state-of-the-art Transformer models. Addresses a critical gap in low-resource South Asian NLP.

Task

Sentiment Analysis

Language

Roman Urdu

Models tested

XLM-RoBERTa, mBERT, and others

Privacy

Privacy-preserving embedding pipeline

DOI

10.21203/rs.3.rs-9827763/v1

Harvard Dataverse HuggingFace Preprint: 10.21203/rs.3.rs-9827763/v1

RUEmoCorp

Harvard Dataverse · HuggingFace

Emotion ClassificationRoman UrduMulti-instituteHarvard Dataverse

The first large-scale Roman Urdu emotion corpus. Multi-institute annotation with substantial inter-annotator agreement (Fleiss κ = 0.658). 134K labeled samples across multiple emotion categories — fully open-source.

Task

Emotion Classification

Language

Roman Urdu

Size

134K labeled samples

Agreement

Fleiss κ = 0.658 (substantial)

Annotation

Multi-institute

DOI

10.21203/rs.3.rs-9759243/v1

Harvard Dataverse HuggingFace Preprint: 10.21203/rs.3.rs-9759243/v1