SINAI/ALIA-es-cultural-heritage-pairs

Dataset Introduction

The ALIA Spanish Cultural and Heritage Retrieval Pairs Corpus contains tabular instances designed to train and evaluate retrieval-oriented models (e.g., dense retrievers / embedding encoders) using passage-grounded query data produced with a Qwen-style prompting workflow integrated in the ALIA encoders pipeline.

It preserves provenance to the original document and passage while exposing controls such as question type and difficulty (ranging from high_school to phd level).

Dataset Description

Dataset Summary

• Data format: One row per passage-based training/evaluation instance.
• Method: query is generated from passage using a Qwen-style LLM prompting approach defined in the project scripts.
• Difficulty scale: difficulty is a categorical label with three levels: high_school, university, or phd.
• Scope: Focused on domain-specific cultural heritage and humanities text, and compatible with document segmentation workflows that track document/chunk provenance.

Dataset Sources

• Project / organization: SINAI Research Group (Universidad de Jaén) and ALIA resources.
• Generation pipeline code: ALIA-UJA encoders scripts (triplets).
• Base corpus: Derived from SINAI/ALIA-es-cultural-heritage heritage corpus integrating 102+ heritage sources.

Uses

• Training dense retrieval encoders from query–passage pairs in cultural heritage domains.
• Building evaluation sets stratified by question type and difficulty for heritage-focused QA systems.
• Domain-specific experiments and ablations using source_id and document/chunk identifiers in cultural heritage contexts.

Dataset Structure

Data Instances

Each row represents a passage-grounded query instance with identifiers for passage and document provenance.

Example (schematic):

{
  "id_passage": "Patrimonio_Cultural_Inmaterial_España-PCI_Nacional_4-10",
  "id_passage_query": "Patrimonio_Cultural_Inmaterial_España-PCI_Nacional_4-10_1",
  "id_document": "PCI_Nacional_4",
  "passage": "El Toque Manual de Campanas, declarado en 2019, es un lenguaje sonoro que ha funcionado a lo largo de los siglos como un medio de comunicación, cumpliendo un conjunto de funciones sociales para la comunidad: informar, coordinar, delimitar el territorio y proteger. Los toques de campanas, basados en el ritmo, han sido los encargados de organizar la vida comunitaria, de delimitar el tiempo y el espacio laboral, diario, festivo y de duelo.",
  "character": "Zapotec ritual expert",
  "type": "background",
  "difficulty": "university",
  "query": "¿Qué funciones sociales y comunicativas tiene el Toque Manual de Campanas declarado en 2019 y de qué manera se está protegiendo este patrimonio inmaterial en España?",
  "source_id": "Patrimonio_Cultural_Inmaterial_España"
}

Data Fields (schema)

Field	Type	Description
id_passage	string	Identifier for the passage unit (document segment).
id_passage_query	string	Unique identifier for the generated query from the passage unit (document segment).
id_document	string	Identifier of the source document (used to group passages).
passage	string	Passage text used as grounding context for query.
character	string	Character who would ask the question.
type	string	Question type/category defined by the generation pipeline.
difficulty	string	Target educational level of the question: one of high_school, university, or phd.
query	string	Question generated/selected from the passage using the Qwen-based prompting methodology.
source_id	string	Provenance identifier (source/domain/corpus) for filtering and auditability.

Data Statistics

• Total instances: 716,131 instances.
• Columns: 9 (id_passage, id_passage_query, id_document, passage, character, type, difficulty, query, source_id).

Distribution by Difficulty:

• university: 394,683
• high_school: 218,829
• phd: 102,594

Distribution by Type (top types):

• background: 216,821
• summary: 158,089
• acquire_knowledge: 143,881
• keywords: 114,737
• interpretation: 49,731
• procedural: 11,278
• requirements: 6,048
• eligibility: 4,740
• consequences: 2,598
• clarification: 2,446
• compliance_check: 2,233
• yes_or_no: 1,085
• deadline: 421
• And 42 additional types including schedule, methodology, description, provenance, timeline, comparison, citation, iconography, materials, location, etc.

Distribution by Source:

• Patrimonio_Cultural_Inmaterial_España: Heritage intangible cultural assets
• Patrimonio_Castilla_Y_Leon: Castilla and León heritage resources
• Wikipedia_Cultura_España: Spanish cultural Wikipedia
• Revistas_Culturales_Biblioteca_Virtual_Prensa_Historica: Cultural journals and historical press
• Revistas_CSIC: CSIC institutional journals
• And 97+ additional heritage sources covering regional inventories, archaeological repositories, museums, archives, and cultural institutions

Example Usage

from datasets import load_dataset

# Replace with your final dataset repo id on Hugging Face
ds = load_dataset("SINAI/ALIA-es-cultural-heritage-pairs", split="train")

print(ds.keys())
print(ds["query"][0])
print(ds["passage"][0][:200])

# Filter by difficulty level
university_level = ds.filter(lambda x: x['difficulty'] == 'university')
phd_level = ds.filter(lambda x: x['difficulty'] == 'phd')

# Filter by question type
background_questions = ds.filter(lambda x: x['type'] == 'background')
heritage_interpretation = ds.filter(lambda x: x['type'] == 'interpretation')

Dataset Creation

Source Data

The dataset is derived from the SINAI/ALIA-es-cultural-heritage heritage corpus, which integrates documentation from 102+ heritage and cultural sources. Documents are segmented into passages, keeping document-level provenance via id_document and passage-level provenance via id_passage.

Data Collection and Processing

• Passages are created via document segmentation from heritage and cultural sources and linked to id_document/id_passage for traceability.
• Questions are produced using a Qwen-driven prompt pipeline in the ALIA encoders scripts, with metadata controls such as type and difficulty.
• source_id supports provenance tracking, source-level audits, and cultural heritage domain attribution.
• The pipeline preserves domain-specific terminology, institutional contexts, and cultural-historical narratives.

Annotations

This dataset is primarily built via automatic or LLM-assisted generation rather than mandatory manual annotation, though downstream filtering/review can be applied depending on the pipeline configuration.

Personal and Sensitive Information

Because the underlying domain is cultural heritage and humanities, instances may include names of historical figures, cultural entities, institutions, place names, and other entities present in the source material. Most data derives from public institutional repositories and academic sources. Additional filtering may be required depending on deployment constraints and cultural sensitivity considerations.

Considerations for Using the Data

Social Impact of Dataset

This dataset can improve access to domain-specific retrieval and QA capabilities in cultural heritage settings by enabling better model adaptation to specialized heritage narratives, historical discourse, institutional documentation, and cultural terminology. It supports democratization of heritage information and enables AI systems to better understand and provide access to cultural resources.

Discussion of Biases

• Domain bias: language style and vocabulary are specialized to cultural heritage, humanities, and institutional documentation and may reduce out-of-domain generalization.
• Source bias: the dataset integrates content from institutional heritage repositories and academic journals, which means models may overfit to formal academic and institutional registers over colloquial heritage narratives.
• Regional and institutional bias: coverage depends on digitization and publication practices across heritage institutions; some regions or cultural domains may be over-represented relative to others.
• Historical and editorial bias: heritage narratives may reflect dominant historiographic, institutional, or curatorial perspectives from the source institutions.
• Difficulty bias: the distribution across high_school, university, and phd levels may not be uniform, potentially skewing model performance toward the most represented level.

Other Known Limitations

• Leakage risk: if splitting is done at passage-level rather than document-level, near-duplicate context may appear across splits; splitting by id_document is recommended.
• LLM artifacts: synthetic questions may contain stylistic patterns that differ from real heritage researcher or user queries, and may require additional filtering for production use.
• Difficulty calibration: the assignment of high_school, university, or phd difficulty levels is model-driven and may not perfectly align with human heritage expert or educator judgments.
• Cultural representation: coverage reflects the breadth of digitized heritage sources; some aspects of cultural heritage (particularly intangible heritage, diaspora heritage, or marginalized cultural traditions) may be under-represented.

Citation

@misc{ALIA-es-cultural-heritage-pairs,
  title={ALIA Spanish Cultural and Heritage Retrieval Pairs Corpus},
  author={SINAI Research Group},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/SINAI/ALIA-es-cultural-heritage-pairs}}
}

---

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ALIA.

Acknowledgments

This dataset has been generated thanks to SCAYLE (Centro de Supercomputación de Castilla y León) which provided the needed computational resources on its CALENDULA supercomputing cluster.

---

Contact: ALIA Project - SINAI Research Group - Universidad de Jaén

More Information: SINAI Research Group | ALIA-UJA Project