Datasets:

yjoonjang
/

mlir-benchmarks

license: mit
language:
  - ar
  - de
  - nl
  - en
  - es
  - fr
  - hi
  - id
  - it
  - ja
  - pt
  - ru
  - vi
  - zh
task_categories:
  - text-retrieval
pretty_name: MLIR Benchmarks
configs:
  - config_name: MLIR-en_Belebele
    default: true
    data_files:
      - split: queries
        path: viewer/MLIR-en/Belebele_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-en/Belebele_test/corpus.jsonl
  - config_name: MLIR-en_MLQA
    data_files:
      - split: queries
        path: viewer/MLIR-en/MLQA_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-en/MLQA_test/corpus.jsonl
  - config_name: MLIR-en_MultiEup
    data_files:
      - split: queries
        path: viewer/MLIR-en/MultiEup_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-en/MultiEup_test/corpus.jsonl
  - config_name: MLIR-en_XQuAD
    data_files:
      - split: queries
        path: viewer/MLIR-en/XQuAD_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-en/XQuAD_test/corpus.jsonl
  - config_name: MLIR-zh_Belebele
    data_files:
      - split: queries
        path: viewer/MLIR-zh/Belebele_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-zh/Belebele_test/corpus.jsonl
  - config_name: MLIR-zh_MLQA
    data_files:
      - split: queries
        path: viewer/MLIR-zh/MLQA_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-zh/MLQA_test/corpus.jsonl
  - config_name: MLIR-zh_XQuAD
    data_files:
      - split: queries
        path: viewer/MLIR-zh/XQuAD_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-zh/XQuAD_test/corpus.jsonl
  - config_name: MLIR-vi_Belebele
    data_files:
      - split: queries
        path: viewer/MLIR-vi/Belebele_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-vi/Belebele_test/corpus.jsonl
  - config_name: MLIR-vi_MLQA
    data_files:
      - split: queries
        path: viewer/MLIR-vi/MLQA_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-vi/MLQA_test/corpus.jsonl
  - config_name: MLIR-vi_XQuAD
    data_files:
      - split: queries
        path: viewer/MLIR-vi/XQuAD_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-vi/XQuAD_test/corpus.jsonl
  - config_name: MLIR-hi_Belebele
    data_files:
      - split: queries
        path: viewer/MLIR-hi/Belebele_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-hi/Belebele_test/corpus.jsonl
  - config_name: MLIR-hi_MLQA
    data_files:
      - split: queries
        path: viewer/MLIR-hi/MLQA_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-hi/MLQA_test/corpus.jsonl
  - config_name: MLIR-hi_XQuAD
    data_files:
      - split: queries
        path: viewer/MLIR-hi/XQuAD_test/queries.jsonl
      - split: corpus
        path: viewer/MLIR-hi/XQuAD_test/corpus.jsonl
  - config_name: MMLIR_Belebele
    data_files:
      - split: queries
        path: viewer/MMLIR/Belebele_test/queries.jsonl
      - split: corpus
        path: viewer/MMLIR/Belebele_test/corpus.jsonl
  - config_name: MMLIR_MLQA
    data_files:
      - split: queries
        path: viewer/MMLIR/MLQA_test/queries.jsonl
      - split: corpus
        path: viewer/MMLIR/MLQA_test/corpus.jsonl
  - config_name: MMLIR_MultiEup
    data_files:
      - split: queries
        path: viewer/MMLIR/MultiEup_test/queries.jsonl
      - split: corpus
        path: viewer/MMLIR/MultiEup_test/corpus.jsonl
  - config_name: MMLIR_XQuAD
    data_files:
      - split: queries
        path: viewer/MMLIR/XQuAD_test/queries.jsonl
      - split: corpus
        path: viewer/MMLIR/XQuAD_test/corpus.jsonl
  - config_name: NeuCLIR_NeuCLIR2022
    data_files:
      - split: queries
        path: viewer/NeuCLIR/NeuCLIR2022_test/queries.jsonl
      - split: corpus
        path: viewer/NeuCLIR/NeuCLIR2022_test/corpus.jsonl
  - config_name: NeuCLIR_NeuCLIR2023
    data_files:
      - split: queries
        path: viewer/NeuCLIR/NeuCLIR2023_test/queries.jsonl
      - split: corpus
        path: viewer/NeuCLIR/NeuCLIR2023_test/corpus.jsonl

MLIR Benchmarks

Multilingual Information Retrieval (MLIR) evaluation benchmarks used in:

MIMO: Multilingual Information Retrieval via Monolingual Objectives (EMNLP 2026)
SHIFT: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval (EMNLP 2026 Findings)

Four multilingual QA/retrieval datasets (Belebele, MLQA, XQuAD-R, MultiEuP-v2) adapted to the MLIR setting: all language versions of the context passages are pooled into one mixed corpus, and for each query the gold context in every language is a positive.

Layout

MLIR-en/   Belebele_test, MLQA_test, XQuAD_test, MultiEup_test   # English source queries (SHIFT main results)
MLIR-zh/   Belebele_test, MLQA_test, XQuAD_test                  # Chinese source queries
MLIR-vi/   ...                                                   # Vietnamese source queries
MLIR-hi/   ...                                                   # Hindi source queries
MMLIR/     Belebele_test, MLQA_test, XQuAD_test, MultiEup_test   # queries in ALL languages (MIMO main results)
NeuCLIR/   NeuCLIR2022_test, NeuCLIR2023_test                    # English queries x mixed ru+zh corpus

The NeuCLIR/ tracks follow the NeuCLIRBench MLIR recipe: the per-language mteb NeuCLIR HardNegatives corpora for Russian and Chinese are pooled (doc ids prefixed rus:: / zho::; Persian excluded), queries are the official English ones from neuclir/bench, and qrels are the union of the cross-lingual qrels.rus ∪ qrels.zho, binarized. The MIMO/SHIFT evaluation code rebuilds these deterministically at runtime; this static copy preserves the exact snapshot used in the papers.

Each benchmark directory contains BEIR-style JSONL files:

queries.jsonl: {"_id": ..., "text": ...[, "lang": ...]}   (MLIR-en/XQuAD uses {qid: text})
corpus.jsonl:  {doc_id: {"text": ..., "lang": ...}}
qrels.jsonl:   {query_id: {doc_id: relevance}}

Note: dt is the mMARCO language code for Dutch (ISO nl).

The viewer/ directory contains record-format mirrors of the same data: {"_id", "text", "lang"} rows for queries/corpus (backing the dataset viewer above) and flat {"query-id", "corpus-id", "score"} rows for qrels. The evaluation code in the MIMO/SHIFT repositories consumes the raw per-benchmark JSONL files.

Usage

These files are consumed directly by the evaluation code in the MIMO and SHIFT repositories (downloaded automatically at evaluation time). Construction scripts are included in both repositories (scripts/prepare_data/build_benchmarks.py).

Citation

@inproceedings{jang2026mimo,
  title     = {{MIMO}: Multilingual Information Retrieval via Monolingual Objectives},
  author    = {Jang, Youngjoon and Hong, Seongtae and Lim, Heuiseok},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}
@inproceedings{jang2026shift,
  title     = {{SHIFT}: Semantic Harmonization via Index-side Feature Transformation for Multilingual Information Retrieval},
  author    = {Jang, Youngjoon and Hong, Seongtae and Moon, Hyeonseok and Lim, Heuiseok},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

Please also cite the original datasets (Belebele, MLQA, XQuAD/LAReQA, MultiEuP-v2) when using these benchmarks.