SINAI/ALIA-es-cultural-heritage-synthetic-instructions

Dataset Introduction

The ALIA Spanish Cultural and Heritage Synthetic Instructions Corpus is a synthetic instruction-tuning resource in Spanish created under the ALIA project using the Magpie methodology. It was designed to train and evaluate language models in cultural heritage, digital humanities, and historical knowledge tasks with natural linguistic variation and large-scale supervision.

It contains:

• 748,480 instances
• 629,682,398 tokens
• 25 task modalities (heritage QA, true/false, topic/genre/period classification, entity extraction, summarization, simplification, and archival analysis)

Dataset Details

Dataset Description

This dataset contains synthetic cultural heritage and history instruction-response pairs generated with instruction models and curated through a multi-step quality pipeline. The generation process follows Magpie, where the model first produces a user-side query (heritage question, historical inquiry, or archival task) and then generates the corresponding answer.

The corpus includes both canonical and naturally varied prompts built from Spanish cultural heritage documents (historical texts, museum catalogs, archival descriptions, academic heritage publications). It includes diverse task types such as heritage QA, cultural classification, entity extraction, text generation, and analysis of historical and patrimonial content, with multiple linguistic style variations (formal, colloquial, exclamatory, telegraphic).

Uses

This dataset is intended for:

• Instruction tuning of cultural heritage LLMs in Spanish
• Heritage and historical question answering
• Digital humanities and archival text analysis
• Cultural entity recognition and classification
• Historical content summarization and simplification
• Cultural knowledge retrieval and digital humanities applications

Dataset Structure

Data Instances

Each instance follows this structure:

{
  "instance_id": "GALTEA--qa_poco_variado-sin_contexto--general_heritage_es_qa_poco_variado-sin_contexto",
  "question": "¿Cuáles son los elementos arquitectónicos más característicos del arte románico español?",
  "response": "Los elementos arquitectónicos más característicos del arte románico español incluyen: arcos de medio punto, bóvedas de cañón, pilares cruciformes, decoración de arquivoltas, portadas monumentales con tímpanos esculpidos, y ornamentación geométrica. La arquitectura románica española refleja influencias de las rutas de peregrinación, especialmente el Camino de Santiago, combinando influencias del norte de Europa con tradiciones locales.",
  "instruction": "Explica los elementos arquitectónicos más característicos del arte románico español.",
  "task_type": "qa_poco_variado",
  "question_variation_style": "original",
  "question_model": "phi-4",
  "response_model": "phi-4",
  "tokens": 298
}

Data Fields

• instance_id (string): Unique identifier of the generated sample.
• question (string): Generated question or instruction prompt presented to the model.
• response (string): Generated answer.
• instruction (string): Full instruction text used for training.
• task_type (string): Generation modality (question/instruction/test format and context settings).
• question_variation_style (string): Variation strategy metadata for the prompt.
• question_model (string): Model used to generate questions/instructions.
• response_model (string): Model used to generate responses.
• tokens (int): Token count of the instance.

Data Splits

Distribution by task_type:

Task Type	Num Instances	Percentage
archival_summarization	11,260	1.50%
argument_analysis	16,028	2.14%
audiovisual_extraction	296	0.04%
audiovisual_qa	15,672	2.09%
comparison	6,856	0.92%
educational_generation	18,792	2.51%
entity_extraction	1,372	0.18%
festive_qa	61,464	8.21%
genre_classification	14,768	1.97%
geographic_classification	6,132	0.82%
metadata_extraction	6,380	0.85%
period_classification	7,752	1.04%
qa	63,292	8.46%
qa_V_F	159,152	21.26%
qa_poco_variado	99,224	13.26%
qa_variado	96,576	12.90%
scientific_qa	63,304	8.46%
scientific_simplification	16,592	2.22%
scientific_summarization	18,172	2.43%
summarization	18,916	2.53%
temporal_extraction	13,088	1.75%
text_generation	10,444	1.40%
text_simplification	17,844	2.38%
topic_classification	2,228	0.30%
typology_classification	2,876	0.38%
TOTAL	748,480	100.00%

Distribution by question_variation_style:

Variation Style	Num Instances	Percentage
original	561,360	75.00%
faltas_ortografia	14,581	1.95%
telegrafico_exclamativo	14,566	1.95%
chat_exclamativo	14,525	1.94%
exclamativo_con_faltas	14,506	1.94%
telegrafico	14,505	1.94%
coloquial_con_faltas	14,482	1.93%
coloquial_dubitativo	14,357	1.92%
dubitativo	14,343	1.92%
dubitativo_con_faltas	14,321	1.91%
exclamativo	14,290	1.91%
coloquial_exclamativo	14,262	1.91%
coloquial	14,253	1.90%
informal_chat	14,129	1.89%

Distribution by question_model:

Model	Num Instances	Percentage
phi-4	561,360	75.00%
gemma-4-31B-it	187,120	25.00%

Distribution by response_model:

Model	Num Instances	Percentage
phi-4	374,240	50.00%
gemma-4-31B-it	187,120	25.00%
Qwen3.6-35B-A3B	135,620	18.12%
kimi-k2.6	51,500	6.88%

Example Usage

from datasets import load_dataset

# Load the complete dataset split
data = load_dataset(
    "SINAI/ALIA-es-cultural-heritage-synthetic-instructions",
    split="train",
)

# Access one example
example = data[0]
print(example["instance_id"])
print(example["task_type"])
print(example["question"][:300])
print(example["response"])

Streaming (recommended for large files):

from datasets import load_dataset

stream_data = load_dataset(
    "SINAI/ALIA-es-cultural-heritage-synthetic-instructions",
    split="train",
    streaming=True,
)

for i, example in enumerate(stream_data):
    print(f"[{i}] {example['task_type']}")
    print(f"Q: {example['question'][:180]}...")
    print(f"A: {example['response'][:180]}...\n")
    if i >= 2:
        break

Dataset Creation

Curation Rationale

This corpus was created to provide large-scale, domain-specialized synthetic supervision for Spanish cultural heritage and digital humanities language tasks within ALIA. The objective is to improve model behavior in heritage QA, cultural analysis, historical reasoning, and archival text processing. The corpus intentionally includes natural linguistic variation (colloquial, telegraphic, exclamatory) to create more naturalistic training data.

Source Data

Context-based generations are grounded on public Spanish cultural and heritage documentation (historical texts, museum records, archival descriptions, academic heritage publications, and institutional repositories) extracted from the ALIA-es-cultural-heritage corpus. General generations are constrained by heritage-domain system prompts trained on cultural knowledge and historical expertise.

Data Collection and Processing

Magpie generation pipeline

The generation process follows Magpie adapted to the cultural heritage domain:

1. A cultural heritage system prompt is provided (trained on historical knowledge, art history, archival practices, and digital humanities expertise).
2. Multiple generation models (Phi-4, Gemma-4-31B-it) complete the user-side prompt by generating heritage questions/instructions (museum documentation, historical analysis, cultural classification, etc.).
3. Multiple response models (Phi-4, Gemma-4-31B-it, Qwen3.6-35B-A3B, Kimi-K2.6) answer the generated queries from different expertise perspectives.
4. Natural linguistic variation is systematically applied: 75% original style, 25% colloquial, telegraphic, exclamatory, and mixed variations to create realistic dialogue.
5. Task types include heritage QA, genre/period/geographic classification, entity extraction, summarization, simplification, argument analysis, and archival description.

This implementation is documented on GitHub in sinai-uja/ALIA-UJA/documentation/data/llms/instructions.

Quality and reconstruction steps

The final corpus reflects the following curation strategy:

1. Initial synthetic generation and diagnostic review.
2. Raw instruction-response outputs produced with Magpie are filtered with Galtea using an LLM-based quality strategy.

The resulting train file is the consolidated artifact used for downstream instruction tuning.

Annotations

No manual annotations are included. Metadata fields are automatically produced during generation and consolidation.

Personal and Sensitive Information

The dataset is generated from public cultural-heritage sources and synthetic transformations. Standard filtering was applied to reduce sensitive or personally identifiable content.

Considerations for Using the Data

Social Impact of Dataset

The dataset can facilitate Spanish cultural heritage AI systems and improve access to cultural knowledge, digital humanities research, and heritage preservation technologies. It supports the development of tools for archival analysis, museum documentation, historical research, and cultural education.

Discussion of Biases

Potential biases include:

• Bias inherited from source cultural heritage discourse, which may reflect institutional and academic perspectives on what is considered "heritage"
• Bias introduced by the multiple generating models used
• Style homogenization typical of synthetic corpora despite linguistic variation attempts
• Potential overrepresentation of well-documented heritage (European, institutional) over community-based or minority heritage
• Language and cultural biases inherent to Spanish cultural terminology and Western heritage frameworks
• Possible underrepresentation of intangible heritage, indigenous knowledge, or contemporary cultural practices
• Geographic bias toward well-known heritage sites and institutions

Other Known Limitations

• Synthetic cultural content may not capture the lived experience and authentic narratives of heritage communities
• Some context snippets may contain OCR or formatting artifacts inherited from digitized heritage documents
• The corpus reflects primarily Spanish institutional heritage perspectives and may differ from community-based or decentralized heritage knowledge
• Task balance emphasizes QA and classification tasks over other heritage-relevant modalities
• Generated text may not reflect the nuance of professional archival or museological practice
• Temporal coverage of heritage varies; contemporary and emerging heritage is underrepresented
• Linguistic variation is synthetic and may not fully capture natural conversational patterns about heritage

Citation

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

---

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