Multi-RAG-Multilingual
Dataset Summary
Multi-RAG-Multilingual extends the RAG_Multilingual dataset, created by the Language Technologies Unit (LangTech) at the Barcelona Supercomputing Center, to six languages: English, Spanish, Catalan, French, German, and Italian.
The original RAG_Multilingual is an instruction-following, synthetic QA dataset built from extractive QA datasets in Catalan, English, and Spanish. This extension adds French, German, and Italian through an automatic translation pipeline.
Original Dataset Characteristics
RAG_Multilingual was created by prompting the Mixtral-8x7B mixture-of-experts model to give more human-like, generative answers to questions based on a given context. The reference datasets were:
- Catalan: CatalanQA
- English: SQuAD
- Spanish: SQAC
The context is typically a short passage from domains such as history, geography, politics, literature, music, and science. A direct quotation from the context serves as a "kernel of truth" that anchors each generated answer. The texts are accessible and do not require advanced domain expertise: they contain limited specialized terminology and mostly straightforward grammar.
The original dataset contains 56,406 instances and can be used to fine-tune models for multilingual Retrieval-Augmented Generation (RAG) systems.
Dataset Creation
To build the extension, the English instances were divided into three equal parts and the Spanish instances into two equal halves. Catalan was left unchanged. The translation model was Qwen3.5-122B-A10B-FP8.
| Source language | Split | Result |
|---|---|---|
| English | 1/3 | Kept as is (en) |
| English | 1/3 | Translated into French (fr) |
| English | 1/3 | Translated into German (de) |
| Spanish | 1/2 | Kept as is (es) |
| Spanish | 1/2 | Translated into Italian (it) |
| Catalan | all | Kept as is (ca) |
Language Distribution
The dataset has six subsets, one per language (42,302 examples in total):
| Language | Subset | Percentage | Examples |
|---|---|---|---|
| English | en |
11.8% | 4,999 |
| Italian | it |
13.3% | 5,632 |
| French | fr |
11.8% | 4,998 |
| Spanish | es |
13.3% | 5,631 |
| German | de |
11.8% | 4,999 |
| Catalan | ca |
37.9% | 16,043 |
Dataset Structure
Fields
| Field | Description |
|---|---|
id |
Unique entry ID in the format {lang}_{num} (e.g., en_0) |
instruction |
The question posed to the model |
context |
A short passage containing the information needed to answer the question |
response |
The answer, based on the context |
category |
Entry type (here, always qa) |
extractive |
A direct quotation from the context that anchors the response |
Example
{
"id": "en_0",
"instruction": "What area is usually grouped with the Appalachians?",
"context": "In addition to the true folded mountains, known as the ridge and valley province, the area of dissected plateau to the north and west of the mountains is usually grouped with the Appalachians. This includes the Catskill Mountains of southeastern New York, the Poconos in Pennsylvania, and the Allegheny Plateau of southwestern New York, western Pennsylvania, eastern Ohio and northern West Virginia. This same plateau is known as the Cumberland Plateau in southern West Virginia, eastern Kentucky, western Virginia, eastern Tennessee, and northern Alabama.",
"response": "The Appalachian region typically includes a dissected plateau to the north and west of the main mountain range. This plateau encompasses several highland areas, such as the Catskill Mountains in New York, the Poconos in Pennsylvania, and the Allegheny Plateau in New York, Pennsylvania, Ohio, and West Virginia. Furthermore, it extends into southern West Virginia, eastern Kentucky, western Virginia, eastern Tennessee, and northern Alabama, where it is known as the Cumberland Plateau.",
"category": "qa",
"extractive": "dissected plateau to the north and west of the mountains"
}
Limitations
The French, German, and Italian entries were translated automatically. They may inherit imperfections from the original dataset and contain translation inconsistencies, mainly involving proper names, acronyms, homographs, and "false friends" within Romance languages.
Several mitigation steps were taken during development, and multiple QA review iterations were performed. Some minor issues may remain.
Licensing
This work is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License, as is the original dataset.
Citation
If you use this dataset, please cite the Multi-SFT paper, which describes the whole SFT collection used to train the Villanova model family and the dataset itself.
@misc{baiamonte2026multisft,
title={Multi-SFT: Permissively-Licensed Multilingual Instruction-Following},
author={Baiamonte, Daniela and Zeinalipour, Kamyar and Fano, Elena and Simonazzi, Stefano and Zugarini, Andrea},
year={2026},
eprint={2610.22086},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2610.22086}
}
- Downloads last month
- 3