Nepali-Nli-20k

---

Nepali Natural Language Inference Dataset

This dataset contains synthetic Natural Language Inference (NLI) examples in the Nepali language. Each example consists of a premise, a hypothesis, and a label indicating the logical relationship between them.

The dataset is designed for training and evaluating Nepali language understanding models, especially sentence embedding models, retrieval models, cross-encoders, and other models that need to understand semantic similarity, contradiction, and entailment in Nepali.

Labels

The dataset follows the standard Natural Language Inference format:

• entailment: The hypothesis logically follows from the premise.
• contradiction: The hypothesis conflicts with the premise.
• neutral: The hypothesis is related to the premise, but cannot be proven true or false from it.

Example

{
  "premise": "दशैंमा धेरै नेपाली परिवारका सदस्यहरू घर फर्केर टीका र जमरा लगाउने गर्छन्।",
  "hypothesis": "दशैंमा परिवारका सदस्यहरू भेटघाट गर्ने चलन हुन्छ।",
  "label": "entailment"
}

Topics Covered

The dataset covers a wide range of Nepali and general real-world topics, including:

• Nepali culture and traditions
• Festivals, rituals, and social customs
• Geography, environment, and tourism
• Education and student life
• Health, family, and daily life
• Banking, insurance, and public services
• Digital payments, mobile apps, and internet usage
• Employment, migration, and workplace situations
• Politics, governance, and civic participation
• Media, misinformation, and digital literacy
• Agriculture, local markets, and small businesses
• Technology, AI, and Nepali language processing

The goal is to make the examples diverse and useful for models that need to understand Nepali in realistic everyday contexts.

Intended Uses

This dataset can be used for:

• Training Nepali sentence embedding models
• Fine-tuning NLI or classification models
• Training cross-encoders for semantic matching
• Building retrieval and reranking systems
• Evaluating Nepali semantic understanding
• Creating triplet datasets for contrastive learning
• Improving Nepali question-answering and search systems

Dataset Creation

The dataset was synthetically generated using a local instruction-following language model. For each row, a topic and NLI label were sampled, and the model generated a Nepali premise and hypothesis pair matching the requested label.

The examples were generated with explicit constraints to ensure:

• Both premise and hypothesis are written in Nepali
• The label relationship is logically consistent
• The topic is reflected in the generated text
• The output follows a strict structured format

Limitations

Because this is a synthetic dataset, some examples may contain:

• Imperfect logical relationships
• Repetitive sentence patterns
• Minor language or grammar issues
• Culturally plausible but not fact-verified statements
• Simplified real-world scenarios

Users are encouraged to inspect, filter, and validate the dataset before using it for high-stakes applications.

Recommended Processing

Before training, it may be useful to:

• Remove duplicate premise-hypothesis pairs
• Validate label consistency with a separate model or human review
• Filter very short or low-quality examples
• Balance the number of examples per label
• Split the dataset into train, validation, and test sets
• Optionally convert the NLI labels into sentence-pair or triplet formats

Citation

If you use this dataset, please cite it as:

@dataset{subedi_nepali_nli_dataset,
  author = {Subedi, Sanjaya},
  title = {Nepali Natural Language Inference Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/jangedoo/nepali-nli-20k}
}

---

🚀 Quick Start

from datasets import load_dataset

# Load the main dataset
dataset = load_dataset("jangedoo/nepali-nli-20k", "data", split="train")
df = dataset.to_pandas()