NLS-CH-Multimodal: A Large-Scale Multi-Modal Cultural Heritage Corpus

Dataset Summary

NLS-CH-Multimodal is a large-scale multimodal corpus derived from the National Library of Scotland (NLS) digital collections, comprising over 512,000 files and exceeding 1 TB in size. The dataset is designed to support Information Retrieval (IR), Retrieval-Augmented Generation (RAG), and analysis of large language model (LLM) behaviour on historical data.

The corpus integrates multiple aligned modalities, including OCR/HTR-derived text, high-resolution document images, page-level layout annotations (ALTO XML), and structured metadata in unified CSV/JSON formats. It spans over five centuries (15th–21st century) and covers domains such as social history, political records, and cartographic collections.

Unlike the original NLS digital archive—designed primarily for human browsing—this dataset restructures materials into machine-readable, IR-ready representations. In addition, it includes a reproducible evaluation benchmark consisting of queries and relevance judgements, enabling systematic experimentation that is not supported by the original archive.

---

Key Contributions

• A large-scale multimodal cultural heritage corpus (>512K files, >1TB)
• Unified representation of text, image, layout (ALTO XML), and metadata
• IR- and RAG-ready structure enabling direct use in retrieval pipelines
• A human-reviewed evaluation benchmark with queries and relevance judgements
• A machine-readable transformation of heterogeneous archival collections

Dataset Structure

The corpus is organised into two primary directory trees:

data/collections/

Each NLS source collection has its own subdirectory organised by temporal period, containing:

• doc_id.txt — full OCR/HTR transcript
• doc_id.pdf — composite scanned PDF
• collection_name.csv — structured metadata (title, author, year, URL, etc.)
• description.txt — collection-level summary

data/collections/

Contains three thematic E-Resources sub-collections:

• Africa_And_New_Imperialism/ — Colonial-era records and court documents
• Indiaraj/ — India-related archival materials
• Slavery/ — Slavery and abolition records

Documents are organised as: year/document_title/doc_id.txt and .pdf Documents with missing year metadata are stored under Unknown/.

---

Data Statistics

The corpus exceeds one terabyte in size and comprises over 512,000 files, spanning the 15th to the 21st century. It includes more than 42,000 OCR-derived plain text documents, 215,567 high-resolution document images, 215,567 page-level OCR representations encoded in ALTO XML, and over 11,000 CSV document-level metadata files, alongside rich metadata expressed in CSV and METS formats. The collection covers broad thematic domains including social history, political records, and extensive cartographic material.

Temporal range: 15th – 21st century Primary languages: English, Scottish Gaelic, Latin, French, Dutch, Portuguese Thematic domains: Social history, political records, cartographic material

---

Intended Use and Supported Tasks

🔍 Information Retrieval

• Benchmarking sparse and dense retrieval on historical corpora
• Cross-era and cross-collection retrieval experiments

🤖 Retrieval-Augmented Generation (RAG)

• End-to-end evaluation of RAG pipelines
• Retrieval grounding for historically complex queries

📊 LLM Evaluation

• Robustness to OCR noise and archaic language
• Behaviour on multilingual and historically situated inputs

🏛️ Digital Humanities

• Longitudinal language analysis
• Named entity recognition and linking
• Document layout and structure analysis

---

Who Can Use This Dataset?

User Group	Use Case
IR researchers	Benchmarking retrieval models on long-tail, historical queries
NLP / LLM researchers	Evaluating models on archaic language and OCR-noisy text
RAG engineers	Testing retrieval pipelines over heterogeneous multimodal corpora
Digital humanities scholars	Longitudinal and cross-cultural document analysis
Cultural heritage institutions	Reproducible baseline for heritage digitisation projects

---

Evaluation Setup

A reproducible evaluation benchmark is included for IR and RAG experiments.

Query Sets

File	Description	Count
evaluation/queries.jsonl	Original factoid query set (Q001–Q044)	44
evaluation/queries_expanded.jsonl	Expanded factoid query set (Q051+)	40
evaluation/queries_all_full.json	Original queries with expected answers and OCR quality scores	44
evaluation/queries_expanded_full.json	Expanded queries with expected answers	40
Total		84 queries

Relevance Judgements (Qrels)

File	Description	Count
evaluation/qrels.tsv	Original relevance judgements	44
evaluation/qrels_expanded.tsv	Expanded relevance judgements	40
evaluation/qrels_verified.tsv	Human-verified relevance judgements	44

All qrels follow the standard TREC format: query-id corpus-id score

Query Schema

Each query record in the .jsonl files follows this structure:

{
  "_id": "Q001",
  "text": "Query text here",
  "metadata": {
    "type": "factoid",
    "difficulty": "easy | medium | hard",
    "status": "verified"
  }
}

The full .json files additionally include expected answers and OCR quality metadata:

{
  "query_id": "Q001",
  "query": "Query text here",
  "query_type": "factoid",
  "expected_answer": "Ground truth answer",
  "difficulty": "medium",
  "doc_id": "22361639",
  "ocr_quality_tier": "high",
  "ocr_quality": 0.985,
  "verified_by": ["Jakub", "yash"],
  "status": "verified"
}

---

Data Collection Methodology

Data was acquired from three NLS platforms:

1. NLS Digital Gallery — OCR transcripts derived from XML/HTM streams
2. Licensed E-Resources — JWT-authenticated downloads via Burp Suite and Postman analysis
3. NLS Data Foundry — Bulk text, image, layout (ALTO XML), spatial data (GeoJSON, KML)

Collection was performed via automated pipelines combining metadata extraction, document retrieval, and validation.

This approach was necessary due to the absence of a unified bulk-access API across all collections.

---

Known Gaps and Limitations

• Post-2000 coverage is sparse — the corpus is primarily pre-21st century
• Single relevance per query — does not yet support multi-document relevance
• OCR noise: historical spelling and scanning artefacts preserved.
• Sensitive content: Period-specific language reflecting historical social norms is present; standard LLM safety filters may flag this content

---

Citation

If you use this dataset, please cite:

@dataset{nls_ch_multimodal_2025,
  author    = {Dhakade, Yash and {NeuraSearch Laboratory}},
  title     = {NLS-CH-Multimodal: A Large-Scale Multi-Modal Cultural
               Heritage Corpus for Information Retrieval},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/NeuraSearchLab/NLS-CH-Multimodal}
}

---

Licence and Attribution

This dataset is derived from materials provided by the National Library of Scotland (NLS) and associated third-party sources.

A sample version of the dataset is publicly available via This HF repository, while the full dataset (exceeding 1 TB) is available upon reasonable request for research purposes. The dataset structure, metadata, and evaluation benchmark (including queries and relevance judgements) are released under the Creative Commons Attribution 4.0 (CC BY 4.0) license. However, the underlying source materials (documents, images, and OCR transcripts) remain subject to the original licensing terms and usage policies of the NLS and any third-party providers.

Users are responsible for complying with the applicable terms of use for each underlying collection and for providing appropriate attribution to the National Library of Scotland and relevant sources.