| --- |
| license: cc-by-4.0 |
| language: |
| - ca |
| - en |
| - es |
| - eu |
| - gl |
| - va |
| pretty_name: IBERtaQA |
| task_categories: |
| - question-answering |
| task_ids: |
| - multiple-choice-qa |
| size_categories: |
| - <1K |
| configs: |
| - config_name: global |
| data_files: |
| - split: ca |
| path: data/BertaQA+.global.ca.jsonl |
| - split: en |
| path: data/BertaQA+.global.en.jsonl |
| - split: es |
| path: data/BertaQA+.global.es.jsonl |
| - split: eu |
| path: data/BertaQA+.global.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.global.gl.jsonl |
| - split: va |
| path: data/BertaQA+.global.va.jsonl |
| - config_name: local-ca |
| data_files: |
| - split: ca |
| path: data/BertaQA+.local-ca.ca.jsonl |
| - split: en |
| path: data/BertaQA+.local-ca.en.jsonl |
| - split: es |
| path: data/BertaQA+.local-ca.es.jsonl |
| - split: eu |
| path: data/BertaQA+.local-ca.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.local-ca.gl.jsonl |
| - split: va |
| path: data/BertaQA+.local-ca.va.jsonl |
| - config_name: local-es |
| data_files: |
| - split: ca |
| path: data/BertaQA+.local-es.ca.jsonl |
| - split: en |
| path: data/BertaQA+.local-es.en.jsonl |
| - split: es |
| path: data/BertaQA+.local-es.es.jsonl |
| - split: eu |
| path: data/BertaQA+.local-es.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.local-es.gl.jsonl |
| - split: va |
| path: data/BertaQA+.local-es.va.jsonl |
| - config_name: local-eu |
| data_files: |
| - split: ca |
| path: data/BertaQA+.local-eu.ca.jsonl |
| - split: en |
| path: data/BertaQA+.local-eu.en.jsonl |
| - split: es |
| path: data/BertaQA+.local-eu.es.jsonl |
| - split: eu |
| path: data/BertaQA+.local-eu.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.local-eu.gl.jsonl |
| - split: va |
| path: data/BertaQA+.local-eu.va.jsonl |
| - config_name: local-gl |
| data_files: |
| - split: ca |
| path: data/BertaQA+.local-gl.ca.jsonl |
| - split: en |
| path: data/BertaQA+.local-gl.en.jsonl |
| - split: es |
| path: data/BertaQA+.local-gl.es.jsonl |
| - split: eu |
| path: data/BertaQA+.local-gl.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.local-gl.gl.jsonl |
| - split: va |
| path: data/BertaQA+.local-gl.va.jsonl |
| - config_name: local-va |
| data_files: |
| - split: ca |
| path: data/BertaQA+.local-va.ca.jsonl |
| - split: en |
| path: data/BertaQA+.local-va.en.jsonl |
| - split: es |
| path: data/BertaQA+.local-va.es.jsonl |
| - split: eu |
| path: data/BertaQA+.local-va.eu.jsonl |
| - split: gl |
| path: data/BertaQA+.local-va.gl.jsonl |
| - split: va |
| path: data/BertaQA+.local-va.va.jsonl |
| --- |
| |
| # Dataset Card for IBERtaQA |
|
|
| [](https://github.com/hitz-zentroa/IBERtaQA) |
|
|
| * **Curated by**: |
| - HiTZ Center, University of the Basque Country (EHU) |
| - Barcelona Supercomputing Center (BSC) |
| - CiTIUS, University of Santiago de Compostela (USC) |
| - GPLSI, University of Alicante (UA) |
| * **Funded by**: Project _Desarrollo de Modelos ALIA_ |
| * **License**: CC-BY-4.0 |
|
|
| ## Dataset Summary |
|
|
| **IBERtaQA** is a multilingual benchmark designed to evaluate the cultural and factual knowledge of language models across Iberian languages |
| and cultural communities. The dataset extends the scope of [BertaQA](https://huggingface.co/datasets/HiTZ/BertaQA) ([Etxaniz et al., 2024](https://proceedings.neurips.cc/paper_files/paper/2024/hash/3bb42f6bb1b1ab6809afd6c90865b087-Abstract-Datasets_and_Benchmarks_Track.html)), |
| a benchmark originally introduced to study how well language models capture both global knowledge and culturally localized knowledge in |
| Basque and English settings. Building on this motivation, **IBERtaQA** broadens the evaluation landscape to multiple Iberian languages and |
| territories. |
|
|
|  |
|
|
| It contains multiple-choice question-answer pairs covering both globally shared knowledge and culturally grounded local knowledge from |
| several Iberian communities. The dataset is organized into six cultural subsets each containing 360 questions balanced across 8 thematic |
| categories and 3 difficulty levels. Questions span domains such as history, literature, geography, music, cinema, sports, traditions, |
| science, and popular culture. All subsets are available in Basque, Catalan, Galician, English, and Valencian through professional human |
| translations, resulting in a fully parallel multilingual benchmark for evaluating cultural knowledge, multilingual robustness, and |
| cross-lingual knowledge transfer. |
|
|
| Unlike many automatically collected trivia datasets, all question-answer pairs in **IBERtaQA** were manually reviewed and curated. Existing |
| questions inherited from BertaQA were verified for factual correctness and updated when necessary. In particular, questions with strong |
| temporal dependencies were rewritten or reformulated to improve the long-term durability of the benchmark. |
|
|
| <ins>**IBERtaQA** is intended exclusively for evaluation and benchmarking purposes and is not intended for model training.</ins> |
|
|
| --- |
|
|
| ## Dataset Details |
|
|
| ### Dataset Structure |
|
|
| **IBERtaQA** is organized using: |
| - **dataset configurations** for cultural subsets: |
| - `global`: Globally shared factual and cultural knowledge |
| - `local-ca`: Catalan cultural and local knowledge |
| - `local-eu`: Basque cultural and local knowledge |
| - `local-es`: Spain-related cultural and local knowledge |
| - `local-gl`: Galician cultural and local knowledge |
| - `local-va`: Valencian Community cultural and local knowledge |
| - and **dataset splits** for language versions: |
| - `eu`: Basque |
| - `en`: English |
| - `es`: Spanish |
| - `ca`: Catalan |
| - `gl`: Galician |
| - `va`: Valencian |
|
|
| Each configuration split contains exactly 360 questions, with all language splits being parallel translations of the same underlying |
| questions. Example usage: |
|
|
| ```python |
| from datasets import load_dataset |
| |
| dataset = load_dataset("HiTZ/IBERtaQA", "global", split="gl") |
| ``` |
|
|
| The example above loads the Galician version of the `global` subset. |
|
|
| ### Categories and Difficulty Levels |
|
|
| Questions are balanced across the following thematic categories: |
|
|
| - Language and Literature |
| - Geography and History |
| - Society and Tradition |
| - Sports and Leisure |
| - Culture and Art |
| - Music and Dance |
| - Science and Technology |
| - Cinema and Shows |
|
|
| And difficulty labels follow a 3-level scale: |
|
|
| 1. Easy |
| 2. Medium |
| 3. Hard |
|
|
| Each set of 360 questions, then, contains 24 cross-section blocks comprising 15 questions. |
|
|
| ### Features |
|
|
| Each example contains the following fields: |
|
|
| | Field | Type | Description | |
| |--------------|----------------|-----------------------------| |
| | `id` | `int` | Unique identifier | |
| | `category` | `string` | Question category | |
| | `group` | `string` | Cultural subset | |
| | `difficulty` | `int` | Difficulty level (1–3) | |
| | `question` | `string` | Question text | |
| | `answer` | `int` | Index of the correct answer | |
| | `candidates` | `list[string]` | Multiple-choice candidates | |
|
|
| The dataset contains only a test split and is intended exclusively for evaluation. |
|
|
| --- |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| **IBERtaQA** is designed as an evaluation-only benchmark for multilingual large language models, culturally grounded language technologies, |
| cross-lingual knowledge transfer, and low-resource language evaluation. |
|
|
| Potential use cases include zero-shot evaluation, multilingual prompting experiments, cultural competence analysis, evaluation of continued |
| pretraining, multilingual alignment studies, and benchmarking of multilingual retrieval-augmented systems. |
|
|
| The benchmark is particularly valuable for studying how language models represent and transfer localized cultural knowledge across languages. |
|
|
| ### Out-of-Scope Use |
|
|
| **IBERtaQA** is NOT intended for: |
| - model training, |
| - knowledge retrieval systems intended for factual guarantees, |
| - or measuring general intelligence beyond cultural and factual multiple-choice reasoning. |
|
|
| As with other benchmark datasets, high performance on **IBERtaQA** should not be interpreted as comprehensive cultural understanding. |
|
|
| --- |
|
|
| ## Evaluation |
|
|
| Performance on **IBERtaQA** is typically measured using exact-match accuracy at subset-level, category-level, difficulty-level accuracy. |
| The multilingual and parallel nature of the benchmark additionally enables evaluation of cross-lingual consistency, language transfer effects, and culturally localized reasoning capabilities. |
| Please find the configuration files to evaluate your models through **Eleuther's LM Harness framework** in the following repository: |
|
|
| * https://github.com/hitz-zentroa/IBERtaQA |
|
|
| --- |
|
|
| ## Dataset Creation |
|
|
| ### Curation Rationale |
|
|
| **IBERtaQA** was created to address the limited representation of Iberian languages and culturally localized knowledge in existing language |
| model evaluation benchmarks. |
|
|
| While many multilingual benchmarks focus primarily on globally dominant languages and general factual knowledge, **IBERtaQA** aims to evaluate |
| whether language models can also capture regional cultures, traditions, historical references, and localized knowledge domains. |
|
|
| The benchmark further seeks to support research on multilingual evaluation, cultural representation, and low-resource language technologies. |
|
|
| ### Data Collection and Processing |
|
|
| The creation of **IBERtaQA** consisted of four main stages. |
|
|
| 1. **Sampling from BertaQA**: |
| To maximize compatibility with the original [BertaQA](https://huggingface.co/datasets/HiTZ/BertaQA) ([Etxaniz et al., 2024](https://proceedings.neurips.cc/paper_files/paper/2024/hash/3bb42f6bb1b1ab6809afd6c90865b087-Abstract-Datasets_and_Benchmarks_Track.html)) |
| benchmark, we first selected 360 questions from the `global` subset, and 360 questions from the `local` Basque cultural subset. |
| The sampling process was designed to maintain a balanced distribution across the 8 thematic categories and the 3 difficulty levels. |
|
|
| 3. **Manual validation and revision**: |
| Expert annotators manually reviewed all sampled question-answer pairs. |
| During this process, we verified factual correctness, corrected outdated or inaccurate information, removed ambiguous formulations, and |
| reformulated highly time-dependent questions. |
| Particular attention was given to improving the long-term durability of the benchmark by avoiding questions whose answers may quickly |
| change over time due to current events or evolving public information. |
|
|
| 5. **Creation of new cultural subsets**: |
| In addition to the original global and Basque subsets, new culturally grounded question sets were manually created for: |
| - Spain (`local-es`), |
| - Catalonia (`local-ca`), |
| - Galicia (`local-gl`), |
| - and Valencian Community (`local-va`). |
|
|
| Each new subset contains 360 questions and was also independently balanced across thematic categories and difficulty levels. The same |
| rationale as in 2. was applied to ensure dataset durability. |
|
|
| 7. **Professional human translation** |
| Finally, all subsets were translated into all supported languages in order to create a fully parallel multilingual benchmark. |
| All translations were produced by professional human translators. |
|
|
| --- |
|
|
| ## Bias, Risks, and Limitations |
|
|
| Although extensive manual validation was performed, several limitations remain: |
|
|
| - The dataset focuses on selected Iberian linguistic communities and does not aim to comprehensively represent all perspectives within |
| those communities. |
| - Translation choices may introduce small linguistic variations across languages. |
| - Benchmark performance may be influenced by models' prior exposure to culturally dominant languages during pretraining. |
|
|
| As with all multiple-choice benchmarks, strong performance may reflect memorization or statistical pattern matching rather than genuine |
| cultural understanding. Users should consider **IBERtaQA** as one component within a broader multilingual evaluation framework. Results |
| should be interpreted carefully, especially when comparing languages with substantially different levels of representation in web-scale |
| training corpora. |
|
|
| --- |
|
|
| ## Funding |
|
|
| This work has been supported and funded by the Ministerio para la Transformación Digital y de la Función Pública and the Plan de |
| Recuperación, Transformación y Resiliencia – funded by the EU through NextGenerationEU within the project _Desarrollo de Modelos ALIA_. |
|
|
| ## Dataset Card Contact |
|
|
| * Naiara Perez Miguel (naiara.perez@ehu.eus) |
| * Paula Rivera Hidalgo de Torralba (paula.rivera@bsc.es) |
| * Iván Martínez-Murillo (ivan.martinezmurillo@ua.es) |
| * Silvia Paniagua Suarez (silvia.paniagua.suarez@usc.es) |
|
|
