sadrmanesh/persian-rag-bench

persian-rag retrieval benchmark

A small Persian (Farsi) retrieval benchmark from the persian-rag project. It is meant for quick, reproducible regression checks of Persian retrieval pipelines, and in particular for how they handle spelling variants. It is too small to rank models in general.

Contents

Config	Rows	Fields
corpus	34	id, title, text
qa	42	id, question, relevant_ids, tags

• corpus: short encyclopedic passages on Persian poets, cities, landmarks, festivals, food, and everyday topics such as rooms, boards of directors, and the stock exchange.
• qa: natural questions or short queries, each with the ids of the relevant corpus passages. 12 queries carry the spelling tag: they use a different but common spelling of a word in the passage, for example طهران for تهران, اطاق for اتاق, or هیات for هیئت.

{"id": "q06", "question": "طهران", "relevant_ids": ["tehran"], "tags": ["spelling"]}

Usage

from datasets import load_dataset

corpus = load_dataset("sadrmanesh/persian-rag-bench", "corpus", split="test")
qa = load_dataset("sadrmanesh/persian-rag-bench", "qa", split="test")

The project's benchmark script runs every retrieval mode with and without query rewriting:

git clone https://github.com/sadrmanesh/persian-rag && cd persian-rag
pip install -e .
python benchmarks/run_benchmark.py

Reference results

From persian-rag 0.3.1 with the offline hashing embedder (dimension 512), chunk size 400, as published in the benchmark docs:

Configuration	Hit@1	Hit@5	MRR@10	nDCG@10	Hit@1 (spelling)
sparse (BM25)	0.881	0.881	0.881	0.879	0.583
sparse + rewrite	0.952	0.976	0.964	0.967	0.833
hybrid	0.952	0.976	0.964	0.967	0.833

Limitations

• Small and hand-written, with one or few relevant passages per query, so scores saturate quickly and differences of one query are visible as about 0.024.
• Passages are short and mostly general knowledge about Iran and Persian culture.
• Text is not pre-normalized: some passages contain half-spaces and some queries use Arabic forms on purpose. Normalize both sides the same way.

License and citation

MIT, like the persian-rag repository. Written by Naser Sadrmanesh.

@misc{sadrmanesh2026persianrag,
  author = {Sadrmanesh, Naser},
  title = {persian-rag: Persian-first retrieval-augmented generation},
  year = {2026},
  howpublished = {\url{https://github.com/sadrmanesh/persian-rag}}
}