NanoBEIR-ps

Dataset Summary

NanoBEIR-ps is a Pashto retrieval benchmark inspired by the NanoBEIR benchmark. It is designed for evaluating semantic search, dense retrieval, sparse retrieval, hybrid search, and Retrieval-Augmented Generation (RAG) systems in the Pashto language.

The dataset provides a lightweight benchmark that can be used to compare embedding models and information retrieval systems without requiring a large-scale corpus.

Supported Tasks

• Dense Retrieval
• Sparse Retrieval (BM25)
• Hybrid Search
• Semantic Search
• Embedding Evaluation
• Retrieval-Augmented Generation (RAG)

Languages

• Pashto (ps)

Dataset Structure

The dataset follows the common BEIR format.

NanoBEIR-ps/
├── corpus
├── queries
└── qrels

corpus

Contains the searchable documents.

Example:

{
  "_id": "doc1",
  "title": "....",
  "text": "...."
}

queries

Contains user search queries.

Example:

{
  "_id": "q1",
  "text": "...."
}

qrels

Contains relevance judgments.

Example:

{
  "query-id": "q1",
  "corpus-id": "doc1",
  "score": 1
}

Intended Uses

NanoBEIR-ps can be used for:

• Evaluating Pashto embedding models
• Benchmarking retrieval systems
• Building RAG pipelines
• Academic research
• Comparing multilingual retrieval models on Pashto

Evaluation Metrics

Common retrieval metrics include:

• nDCG@10
• Recall@10
• Recall@100
• MRR
• MAP

Limitations

• This is a lightweight benchmark.
• It should not be considered a complete representation of all Pashto domains.
• Retrieval performance may vary across different topics.

Citation

If you use this dataset, please cite:

@dataset{nassim2026_nanobeir_ps,
  author = {Nasibullah Nassim},
  title = {NanoBEIR-ps: A Lightweight Pashto Retrieval Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/nassimjp/NanoBEIR-ps}
}

License

Apache-2.0