MuPLeR-retrieval

An MTEB dataset

Massive Text Embedding Benchmark

MuPLeR: Multilingual Parallel Legal Retrieval

MuPLeR-retrieval is a multilingual, parallel legal dataset designed for evaluating retrieval and cross-lingual retrieval tasks. Dataset contains 10,000 human-translated parallel passages (derived from the European Union's DGT-Acquis corpus) & 200 parallel queries (synthetic) across 14 European languages.

To note: Technical paper writing in process

Task category	Retrieval (text-to-text)
Domains	Legal
Passage data Reference	An Overview of the {{European Union}}'s Highly Multilingual Parallel Corpora

Source datasets:

• eherra/MuPLeR-retrieval

How to evaluate on this task

You can evaluate an embedding model on this dataset using the following code:

import mteb

task = mteb.get_task("MuPLeR-retrieval")
evaluator = mteb.MTEB([task])

model = mteb.get_model(YOUR_MODEL)
evaluator.run(model)

To learn more about how to run models on mteb task check out the GitHub repository.

Dataset Overview

• Languages (14): en, fr, sl, it, pl, fi, nl, pt, sv, sk, lv, lt, el, es.
• Domain: Legal (European Union legislation).
• Passages: 10,000 human-translated parallel legal passages (60–150 words each) per language.
• Queries: 200 synthetic parallel queries per language.
• Source: Derived from the European Commission’s DGT-Acquis corpus.

Dataset Construction

This dataset was constructed by processing the DGT-Acquis Paragraph-Level Corpus (da1-pc) through a custom pipeline designed for high-quality embedding evaluation:

1. Text File Extraction: Paragraph-level text files were downloaded from the DGT-Acquis and merged into a structured format.
2. Text Cleaning: Concatenated words were corrected, and specific leftover formatting tags (e.g., <HT TYPE="SUP">) were removed.
3. Sentence Splitting & Chunking: Texts parts were broken down into strict passages of 60–150 words, with a maximum of 5 sentences per chunk. This short-form structure ensures that models with limited context windows can be evaluated without information loss due to truncation.
4. Deduplication: A 98% fuzzy match threshold was applied to filter out redundant content. This removed near-duplicate passages differing only by minor punctuation, capitalization, or extra spacing.
5. Language Detection: Texts were validated using the lingua-py library to ensure they were in the correct language.
6. Cosine Similarity Alignment: A semantic similarity check (cosine score ≥ 0.75) was performed to verify the cross-lingual alignment of all 14 parallel passages.
7. Final Selection: From the translations that met all the criteria above, the first 10,000 passages were chosen for the final dataset.
8. Synthetic Query Generation: 220 English passages were first selected to generate search questions using LlamaIndex and GPT-5-mini. To ensure parallel consistency, the corresponding passages in the other 13 languages were used to generate matching queries in their respective languages. From this set, the final 200 query-passage pairs per language were selected for the benchmark.
9. Cross-Lingual Query Validation: Each generated query was validated to ensure it was in the correct target language. To ensure high parallel consistency, a final similarity check (cosine score ≥ 0.80) was performed to confirm that the queries remained semantically aligned across all 14 languages.

Final Dataset Composition:

• 14x 10,000 parallel context chunks across the languages
• 14x 200 synthetic questions across the languages, mapped to the corresponding passage.

Source Data

This dataset passages are based on the DGT-Acquis corpus (website, & research paper), which is owned by the European Commission and released under the European Union Public License (EUPL).

• Original Corpus: DGT-Acquis (paragraph-level, 2004–2011).
• Producers: The European Commission, Directorate-General for Translation (DGT).
• Data Processing: The dataset was processed by a custom pipeline, which included text extraction, cleaning, chunk selection, language detection, similarity checks, and synthetic question generation.

Citation

If you use this dataset, please cite the dataset as well as mteb, as this dataset likely includes additional processing as a part of the MMTEB Contribution.

@article{steinberger2014overview,
  author = {Steinberger, Ralf and Ebrahim, Mohamed and Poulis, Alexandros and {Carrasco-Benitez}, Manuel and Schl{\"u}ter, Patrick and Przybyszewski, Marek and Gilbro, Signe},
  doi = {10.1007/s10579-014-9277-0},
  issn = {1574-0218},
  journal = {Language Resources and Evaluation},
  keywords = {DCEP,DGT-Acquis,DGT-TM,EAC-TM,ECDC-TM,Eur-Lex,European Union,EuroVoc,Highly multilingual,JRC EuroVoc Indexer JEX,JRC-Acquis,Linguistic resources,Parallel corpora,Translation memory},
  langid = {english},
  language = {en},
  month = dec,
  number = {4},
  pages = {679--707},
  title = {An Overview of the {{European Union}}'s Highly Multilingual Parallel Corpora},
  urldate = {2026-03-29},
  volume = {48},
  year = {2014},
}


@article{enevoldsen2025mmtebmassivemultilingualtext,
  title={MMTEB: Massive Multilingual Text Embedding Benchmark},
  author={Kenneth Enevoldsen and Isaac Chung and Imene Kerboua and Márton Kardos and Ashwin Mathur and David Stap and Jay Gala and Wissam Siblini and Dominik Krzemiński and Genta Indra Winata and Saba Sturua and Saiteja Utpala and Mathieu Ciancone and Marion Schaeffer and Gabriel Sequeira and Diganta Misra and Shreeya Dhakal and Jonathan Rystrøm and Roman Solomatin and Ömer Çağatan and Akash Kundu and Martin Bernstorff and Shitao Xiao and Akshita Sukhlecha and Bhavish Pahwa and Rafał Poświata and Kranthi Kiran GV and Shawon Ashraf and Daniel Auras and Björn Plüster and Jan Philipp Harries and Loïc Magne and Isabelle Mohr and Mariya Hendriksen and Dawei Zhu and Hippolyte Gisserot-Boukhlef and Tom Aarsen and Jan Kostkan and Konrad Wojtasik and Taemin Lee and Marek Šuppa and Crystina Zhang and Roberta Rocca and Mohammed Hamdy and Andrianos Michail and John Yang and Manuel Faysse and Aleksei Vatolin and Nandan Thakur and Manan Dey and Dipam Vasani and Pranjal Chitale and Simone Tedeschi and Nguyen Tai and Artem Snegirev and Michael Günther and Mengzhou Xia and Weijia Shi and Xing Han Lù and Jordan Clive and Gayatri Krishnakumar and Anna Maksimova and Silvan Wehrli and Maria Tikhonova and Henil Panchal and Aleksandr Abramov and Malte Ostendorff and Zheng Liu and Simon Clematide and Lester James Miranda and Alena Fenogenova and Guangyu Song and Ruqiya Bin Safi and Wen-Ding Li and Alessia Borghini and Federico Cassano and Hongjin Su and Jimmy Lin and Howard Yen and Lasse Hansen and Sara Hooker and Chenghao Xiao and Vaibhav Adlakha and Orion Weller and Siva Reddy and Niklas Muennighoff},
  publisher = {arXiv},
  journal={arXiv preprint arXiv:2502.13595},
  year={2025},
  url={https://arxiv.org/abs/2502.13595},
  doi = {10.48550/arXiv.2502.13595},
}

@article{muennighoff2022mteb,
  author = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Loïc and Reimers, Nils},
  title = {MTEB: Massive Text Embedding Benchmark},
  publisher = {arXiv},
  journal={arXiv preprint arXiv:2210.07316},
  year = {2022}
  url = {https://arxiv.org/abs/2210.07316},
  doi = {10.48550/ARXIV.2210.07316},
}

Dataset Statistics

Dataset Statistics

The following code contains the descriptive statistics from the task. These can also be obtained using:

import mteb

task = mteb.get_task("MuPLeR-retrieval")

desc_stats = task.metadata.descriptive_stats

{
    "test": {
        "num_samples": 142800,
        "number_of_characters": 95547557,
        "documents_text_statistics": {
            "total_text_length": 95151535,
            "min_text_length": 331,
            "average_text_length": 679.6538214285714,
            "max_text_length": 1184,
            "unique_texts": 140000
        },
        "documents_image_statistics": null,
        "documents_audio_statistics": null,
        "queries_text_statistics": {
            "total_text_length": 396022,
            "min_text_length": 54,
            "average_text_length": 141.43642857142856,
            "max_text_length": 234,
            "unique_texts": 2800
        },
        "queries_image_statistics": null,
        "queries_audio_statistics": null,
        "relevant_docs_statistics": {
            "num_relevant_docs": 2800,
            "min_relevant_docs_per_query": 1,
            "average_relevant_docs_per_query": 1.0,
            "max_relevant_docs_per_query": 1,
            "unique_relevant_docs": 2800
        },
        "top_ranked_statistics": null
    }
}