omarelba/letzcross-benchmark

LëtzCross Benchmark

LëtzCross is a cross-lingual page-level retrieval benchmark built from Luxembourgish PDF documents. It evaluates retrieval systems that must identify the relevant page for questions in English, French, German, and Luxembourgish, covering both text-focused and visually grounded retrieval needs.

Dataset structure

Configuration	Rows	Description
corpus	908	Page-level retrieval corpus with extracted text and page images
queries	2,316	579 questions in four languages (EN, FR, DE, LB)
qrels	579	Query-to-page relevance judgments
docs	140	Source-document metadata

Fields

corpus

• corpus_id: Unique page identifier, such as corpus_735
• corpus: Page image stored as a Hugging Face Image feature
• doc_id: Source-document identifier
• text: Extracted page content using layout detection and Luxembourgish-aware OCR.
• image_filename: Page-image filename
• page_number: Page number within the source document

queries

• query_id: Question identifier shared by its four translations, such as q_1
• query: Question text in the language specified by language
• language: Query language (en, fr, de, or lb)
• answers: Annotated answer strings
• final_answer: Final normalized answer
• qa_type: Question/answer category

Every underlying question contributes four rows, one for each language. Its translations intentionally share the same query_id, matching the multilingual experiment setup. Filter by language before creating a query-ID lookup.

qrels

• query_id: Query identifier shared by all translations of the question
• corpus_id: Relevant corpus-page identifier
• score: Relevance score

Qrels are published directly from the benchmark relevance file. Each judgment applies to all four translations that share its query_id.

docs

• doc_id: Unique source-document identifier
• file_name: Source PDF filename
• has_diagram: Whether the document contains diagrams
• has_image: Whether the document contains images
• has_table: Whether the document contains tables
• language: Document language
• num_pages: Number of pages
• total_qa_pairs: Number of associated question-answer pairs

Loading the dataset

The benchmark is available through four Hugging Face configurations, each with a single test split and its own schema.

from datasets import load_dataset

corpus = load_dataset(
    "omarelba/letzcross-benchmark",
    "corpus",
    split="test",
)
queries = load_dataset(
    "omarelba/letzcross-benchmark",
    "queries",
    split="test",
)
qrels = load_dataset(
    "omarelba/letzcross-benchmark",
    "qrels",
    split="test",
)
docs = load_dataset(
    "omarelba/letzcross-benchmark",
    "docs",
    split="test",
)

License

The LëtzCross benchmark data and annotations are released under the Open Data Commons Attribution License (ODC-By) v1.0, following the licensing approach used by FinePDFs.

The original PDF documents remain subject to the licenses and terms of their respective sources and are not relicensed by LëtzCross.

Citation

If you use this dataset, please cite:

LëtzCross: A Cross-Lingual Page-Level Benchmark for Multimodal Retrieval over Luxembourgish Documents

https://arxiv.org/abs/2608.21714

LëtzCross is a collaborative work by:

• Omar El Bachyr — University of Luxembourg (Corresponding author)
• Fred Philippy — University of Luxembourg
• Laura Maria Bernardy — University of Luxembourg
• Saad Ezzini — King Fahd University of Petroleum and Minerals
• Jacques Klein — University of Luxembourg
• Tegawendé F. Bissyandé — University of Luxembourg