RAGBench_Kazakh

Summary

RAGBench_Kazakh is a machine-translated Kazakh version of the original RAGBench benchmark. It is designed to evaluate retrieval-augmented generation (RAG) systems, focusing on how well models use retrieved context to produce grounded answers.

The dataset is built from the test splits of multiple RAGBench subsets covering domains such as biomedical research, general knowledge, legal documents, customer support, and finance. Each example includes a question, retrieved documents, and a reference answer translated into Kazakh, while preserving the original structure for cross-lingual comparison.

All samples come from test splits only. Each example may include multiple retrieved documents, and the batch_key field allows grouping samples by their original subset for more detailed evaluation.

Dataset Structure and Subset Coverage

The dataset is a single merged test split combining all 12 RAGBench subsets. In total, it contains 11,431 examples. Each sample includes a question, a set of retrieved documents, and a reference answer, along with an identifier indicating its original subset.

Subset	Description	Number of examples
covidqa	Biomedical QA focused on COVID-19 research	244
cuad	Legal QA over contract documents	322
delucionqa	Customer-support and hallucination-focused QA	184
emanual	QA over manuals and instructional content	132
expertqa	Expert-level fact-based questions	167
finqa	Financial QA with numerical reasoning	2,144
hagrid	Generative retrieval and general knowledge QA	1,287
hotpotqa	Multi-hop reasoning over multiple documents	374
msmarco	Open-domain web search QA	413
pubmedqa	Biomedical QA over scientific abstracts	2,438
tatqa	Financial QA with mixed text and table reasoning	3,258
techqa	Technical and IT support QA	126

These subsets correspond to broader domains such as biomedical research, general knowledge, legal, customer support, and finance.

Data Fields

Field	Description
batch_key	Unique identifier of the sample, also encoding the source subset (e.g., covidqa__1421).
question	Question text in Kazakh.
documents	Retrieved supporting documents in Kazakh, stored as a serialized list of strings.
response	Reference answer in Kazakh.

Example

{
  "batch_key": "covidqa__1421",
  "question": "Антивирустық тазартудың күшті индукциясының арқасында ұзаққа созылған қабынуды тудырмауы мүмкін вирустар?",
  "documents": [
    "Title: ...",
    "Title: ...",
    "Title: ..."
  ],
  "response": "Антивирустық тазартудың күшті индукциясының арқасында ұзаққа созылған қабынуды тудырмауы мүмкін вирустар - ..."
}