mteb/turkish-rag-eval

Turkish RAG Eval

58 hand-written Turkish questions over 54 Turkish Wikipedia health articles (1.09 M characters), each labelled with the article that answers it and a verbatim answer span. Built for turkish-rag-eval, a harness that measures which parts of a RAG pipeline (chunking, stemming, embedding model) pay off on Turkish.

from datasets import load_dataset

corpus = load_dataset("RizgarOzan/turkish-rag-eval", "corpus", split="corpus")
queries = load_dataset("RizgarOzan/turkish-rag-eval", "queries", split="queries")
qrels = load_dataset("RizgarOzan/turkish-rag-eval", split="test")

Files

BEIR layout, the one MTEB retrieval tasks read.

File	Rows	Fields
corpus.jsonl	54	_id (MediaWiki pageid), title, text, url
queries.jsonl	58	_id, text, answer_span
qrels/test.jsonl	58	query-id, corpus-id, score (always 1)

27 of the 54 articles answer at least one question; the other 27 are distractors from the same domain.

passages/ holds the same data at passage level (configs passages-corpus, passages-queries, passages-qrels): the harness's hierarchical chunks (≤700 characters, title is the section path), and a passage is relevant when it comes from the answering article and contains answer_span. Articles average ~20 000 characters, so a 512-token encoder only sees the lead of each one at article level; passages are what a RAG pipeline actually retrieves.

Full set: 296 questions, six domains

full/ (configs prefixed full-, same layout, article and passage level) adds 238 LLM-drafted questions over 40 more articles in five domains: history, geography, astronomy, biology and computing. 94 articles, 1.89 M characters, 296 questions. Each query has a source field, human or llm-draft, so you can filter either way.

queries = load_dataset("RizgarOzan/turkish-rag-eval", "full-queries", split="queries")
drafted = queries.filter(lambda q: q["source"] == "llm-draft")

The drafts were double-labelled, not human-verified: one model pass wrote the question and marked the span, a second, separate pass marked the span again without seeing the first. Only items where the two passes agree are here (4 disagreements are left out). Agreement figures per batch are in the repo README. Treat the numbers you get on them as provisional.

How it was made

• Questions are paraphrased, not copied. "Şeker hastalığı teşhisi konan kişilerin ne kadarında ketoasidoz da bulunuyor?" is asked of text reading "yaklaşık %25'i, diyabet teşhisi konulduğunda...". Copied wording would hand lexical retrievers an unearned advantage.
• One human annotator, one pass. No inter-annotator agreement figure yet.
• Relevance here is article-level. The harness is stricter: a retrieved chunk counts only if it comes from the right article and contains answer_span. Use the span if you chunk the corpus yourself.
• Corpus snapshot. Articles were fetched through the MediaWiki API as plain-text extracts, == Section == markers kept. The GitHub repo refetches live revisions; this upload pins one snapshot.

Results on this data

From the harness: hierarchical chunks, chunk-level relevance, CPU only. All configurations and six models are in the repo README.

Retriever	nDCG@10
dense, newmindai/Mursit-Large-TR-Retrieval	0.781
dense, intfloat/multilingual-e5-base	0.668
dense, paraphrase-multilingual-MiniLM-L12-v2	0.501
BM25, 5-character prefix stemming	0.494

On the full 296-question set the order changes, which is why the drafts are kept apart: e5-base 0.646, Mursit 0.613, e5-small 0.568, MiniLM 0.405, and stemmed BM25 0.558 (details).

Limits

• 58 queries is small: differences under about 0.05 nDCG are noise.
• The 58 human questions cover one domain (health). The other five domains exist only as LLM drafts in full/ until people check them.
• Encyclopaedic text, not clinical text. No patient data is used anywhere, and nothing here is a medical device.

Licence

Article text and answer spans come from Turkish Wikipedia, CC BY-SA 4.0; each article keeps its url. The questions were written for this dataset and are released under the same licence. The harness code is MIT.

Citation

@misc{ozan2026turkishrageval,
  author       = {Rızgar Ozan},
  title        = {Turkish RAG Eval: a retrieval test set for Turkish Wikipedia},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/RizgarOzan/turkish-rag-eval}},
  note         = {Harness: \url{https://github.com/RizgarOzan/turkish-rag-eval}}
}