metadata
dataset_info:
- config_name: ar_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 1217782171
num_examples: 2061414
download_size: 547328442
dataset_size: 1217782171
- config_name: ar_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 448812
num_examples: 2896
- name: testB
num_bytes: 114733
num_examples: 1405
- name: train
num_bytes: 464511
num_examples: 3495
- name: testA
num_bytes: 71374
num_examples: 936
download_size: 680380
dataset_size: 1099430
- config_name: bn_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 309278144
num_examples: 297265
download_size: 111041827
dataset_size: 309278144
- config_name: bn_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 92875
num_examples: 411
- name: testB
num_bytes: 151662
num_examples: 1130
- name: train
num_bytes: 369128
num_examples: 1631
- name: testA
num_bytes: 16177
num_examples: 102
download_size: 313614
dataset_size: 629842
- config_name: de_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 7018142949
num_examples: 15866222
download_size: 4154850993
dataset_size: 7018142949
- config_name: de_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 47042
num_examples: 305
- name: testB
num_bytes: 50010
num_examples: 712
download_size: 67465
dataset_size: 97052
- config_name: en_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 13936141152
num_examples: 32893221
download_size: 7975410434
dataset_size: 13936141152
- config_name: en_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 114768
num_examples: 799
- name: testB
num_bytes: 134852
num_examples: 1790
- name: train
num_bytes: 403838
num_examples: 2863
- name: testA
num_bytes: 44411
num_examples: 734
download_size: 495734
dataset_size: 697869
- config_name: es_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 4490445463
num_examples: 10373953
download_size: 2569985970
dataset_size: 4490445463
- config_name: es_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 100232
num_examples: 648
- name: testB
num_bytes: 113567
num_examples: 1515
- name: train
num_bytes: 334005
num_examples: 2162
download_size: 381688
dataset_size: 547804
- config_name: fa_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 1091928603
num_examples: 2207172
download_size: 474245644
dataset_size: 1091928603
- config_name: fa_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 115738
num_examples: 632
- name: testB
num_bytes: 145582
num_examples: 1476
- name: train
num_bytes: 384820
num_examples: 2107
download_size: 398966
dataset_size: 646140
- config_name: fi_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 739208952
num_examples: 1883509
download_size: 438735297
dataset_size: 739208952
- config_name: fi_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 171970
num_examples: 1271
- name: testB
num_bytes: 43425
num_examples: 711
- name: train
num_bytes: 332978
num_examples: 2897
- name: testA
num_bytes: 62703
num_examples: 1060
download_size: 415814
dataset_size: 611076
- config_name: fr_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 5507181331
num_examples: 14636953
download_size: 3090892618
dataset_size: 5507181331
- config_name: fr_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 51609
num_examples: 343
- name: testB
num_bytes: 56549
num_examples: 801
- name: train
num_bytes: 171882
num_examples: 1143
download_size: 192265
dataset_size: 280040
- config_name: hi_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 490803869
num_examples: 506264
download_size: 175778700
dataset_size: 490803869
- config_name: hi_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 84874
num_examples: 350
- name: testB
num_bytes: 132031
num_examples: 819
- name: train
num_bytes: 279651
num_examples: 1169
download_size: 258373
dataset_size: 496556
- config_name: id_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 537334548
num_examples: 1446315
download_size: 271117969
dataset_size: 537334548
- config_name: id_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 132827
num_examples: 960
- name: testB
num_bytes: 33388
num_examples: 611
- name: train
num_bytes: 562782
num_examples: 4071
- name: testA
num_bytes: 44066
num_examples: 731
download_size: 505333
dataset_size: 773063
- config_name: ja_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 2970531665
num_examples: 6953614
download_size: 1667379078
dataset_size: 2970531665
- config_name: ja_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 128891
num_examples: 860
- name: testB
num_bytes: 82234
num_examples: 1141
- name: train
num_bytes: 523962
num_examples: 3477
- name: testA
num_bytes: 47589
num_examples: 650
download_size: 517063
dataset_size: 782676
- config_name: ko_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 633206834
num_examples: 1486752
download_size: 361358053
dataset_size: 633206834
- config_name: ko_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 40035
num_examples: 213
- name: testB
num_bytes: 94731
num_examples: 1417
- name: train
num_bytes: 168113
num_examples: 868
- name: testA
num_bytes: 20398
num_examples: 263
download_size: 204713
dataset_size: 323277
- config_name: ru_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 5921641619
num_examples: 9543918
download_size: 2768129481
dataset_size: 5921641619
- config_name: ru_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 230853
num_examples: 1252
- name: testB
num_bytes: 63696
num_examples: 718
- name: train
num_bytes: 743296
num_examples: 4683
- name: testA
num_bytes: 95927
num_examples: 911
download_size: 709053
dataset_size: 1133772
- config_name: sw_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 32766816
num_examples: 131924
download_size: 16590219
dataset_size: 32766816
- config_name: sw_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 68949
num_examples: 482
- name: testB
num_bytes: 29384
num_examples: 465
- name: train
num_bytes: 186466
num_examples: 1901
- name: testA
num_bytes: 39535
num_examples: 638
download_size: 184382
dataset_size: 324334
- config_name: te_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 549992531
num_examples: 518079
download_size: 149673264
dataset_size: 549992531
- config_name: te_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 114574
num_examples: 828
- name: testB
num_bytes: 93547
num_examples: 793
- name: train
num_bytes: 573012
num_examples: 3452
- name: testA
num_bytes: 72265
num_examples: 594
download_size: 366102
dataset_size: 853398
- config_name: th_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 549881577
num_examples: 542166
download_size: 201784561
dataset_size: 549881577
- config_name: th_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 165145
num_examples: 733
- name: testB
num_bytes: 80764
num_examples: 650
- name: train
num_bytes: 591623
num_examples: 2972
- name: testA
num_bytes: 140415
num_examples: 992
download_size: 491456
dataset_size: 977947
- config_name: yo_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 10639617
num_examples: 49043
download_size: 5157435
dataset_size: 10639617
- config_name: yo_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 17153
num_examples: 119
- name: testB
num_bytes: 18321
num_examples: 288
download_size: 24895
dataset_size: 35474
- config_name: zh_corpus
features:
- name: docid
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 1718428423
num_examples: 4934368
download_size: 1092328396
dataset_size: 1718428423
- config_name: zh_queries
features:
- name: query_id
dtype: string
- name: query
dtype: string
- name: positives
sequence: int64
- name: negatives
sequence: int64
splits:
- name: dev
num_bytes: 53829
num_examples: 393
- name: testB
num_bytes: 52592
num_examples: 920
- name: train
num_bytes: 180351
num_examples: 1312
download_size: 197670
dataset_size: 286772
configs:
- config_name: ar_corpus
data_files:
- split: train
path: ar_corpus/train-*
- config_name: ar_queries
data_files:
- split: dev
path: ar_queries/dev-*
- split: testB
path: ar_queries/testB-*
- split: train
path: ar_queries/train-*
- split: testA
path: ar_queries/testA-*
- config_name: bn_corpus
data_files:
- split: train
path: bn_corpus/train-*
- config_name: bn_queries
data_files:
- split: dev
path: bn_queries/dev-*
- split: testB
path: bn_queries/testB-*
- split: train
path: bn_queries/train-*
- split: testA
path: bn_queries/testA-*
- config_name: de_corpus
data_files:
- split: train
path: de_corpus/train-*
- config_name: de_queries
data_files:
- split: dev
path: de_queries/dev-*
- split: testB
path: de_queries/testB-*
- config_name: en_corpus
data_files:
- split: train
path: en_corpus/train-*
- config_name: en_queries
data_files:
- split: dev
path: en_queries/dev-*
- split: testB
path: en_queries/testB-*
- split: train
path: en_queries/train-*
- split: testA
path: en_queries/testA-*
- config_name: es_corpus
data_files:
- split: train
path: es_corpus/train-*
- config_name: es_queries
data_files:
- split: dev
path: es_queries/dev-*
- split: testB
path: es_queries/testB-*
- split: train
path: es_queries/train-*
- config_name: fa_corpus
data_files:
- split: train
path: fa_corpus/train-*
- config_name: fa_queries
data_files:
- split: dev
path: fa_queries/dev-*
- split: testB
path: fa_queries/testB-*
- split: train
path: fa_queries/train-*
- config_name: fi_corpus
data_files:
- split: train
path: fi_corpus/train-*
- config_name: fi_queries
data_files:
- split: dev
path: fi_queries/dev-*
- split: testB
path: fi_queries/testB-*
- split: train
path: fi_queries/train-*
- split: testA
path: fi_queries/testA-*
- config_name: fr_corpus
data_files:
- split: train
path: fr_corpus/train-*
- config_name: fr_queries
data_files:
- split: dev
path: fr_queries/dev-*
- split: testB
path: fr_queries/testB-*
- split: train
path: fr_queries/train-*
- config_name: hi_corpus
data_files:
- split: train
path: hi_corpus/train-*
- config_name: hi_queries
data_files:
- split: dev
path: hi_queries/dev-*
- split: testB
path: hi_queries/testB-*
- split: train
path: hi_queries/train-*
- config_name: id_corpus
data_files:
- split: train
path: id_corpus/train-*
- config_name: id_queries
data_files:
- split: dev
path: id_queries/dev-*
- split: testB
path: id_queries/testB-*
- split: train
path: id_queries/train-*
- split: testA
path: id_queries/testA-*
- config_name: ja_corpus
data_files:
- split: train
path: ja_corpus/train-*
- config_name: ja_queries
data_files:
- split: dev
path: ja_queries/dev-*
- split: testB
path: ja_queries/testB-*
- split: train
path: ja_queries/train-*
- split: testA
path: ja_queries/testA-*
- config_name: ko_corpus
data_files:
- split: train
path: ko_corpus/train-*
- config_name: ko_queries
data_files:
- split: dev
path: ko_queries/dev-*
- split: testB
path: ko_queries/testB-*
- split: train
path: ko_queries/train-*
- split: testA
path: ko_queries/testA-*
- config_name: ru_corpus
data_files:
- split: train
path: ru_corpus/train-*
- config_name: ru_queries
data_files:
- split: dev
path: ru_queries/dev-*
- split: testB
path: ru_queries/testB-*
- split: train
path: ru_queries/train-*
- split: testA
path: ru_queries/testA-*
- config_name: sw_corpus
data_files:
- split: train
path: sw_corpus/train-*
- config_name: sw_queries
data_files:
- split: dev
path: sw_queries/dev-*
- split: testB
path: sw_queries/testB-*
- split: train
path: sw_queries/train-*
- split: testA
path: sw_queries/testA-*
- config_name: te_corpus
data_files:
- split: train
path: te_corpus/train-*
- config_name: te_queries
data_files:
- split: dev
path: te_queries/dev-*
- split: testB
path: te_queries/testB-*
- split: train
path: te_queries/train-*
- split: testA
path: te_queries/testA-*
- config_name: th_corpus
data_files:
- split: train
path: th_corpus/train-*
- config_name: th_queries
data_files:
- split: dev
path: th_queries/dev-*
- split: testB
path: th_queries/testB-*
- split: train
path: th_queries/train-*
- split: testA
path: th_queries/testA-*
- config_name: yo_corpus
data_files:
- split: train
path: yo_corpus/train-*
- config_name: yo_queries
data_files:
- split: dev
path: yo_queries/dev-*
- split: testB
path: yo_queries/testB-*
- config_name: zh_corpus
data_files:
- split: train
path: zh_corpus/train-*
- config_name: zh_queries
data_files:
- split: dev
path: zh_queries/dev-*
- split: testB
path: zh_queries/testB-*
- split: train
path: zh_queries/train-*
license: apache-2.0
MIRACL Unified Dataset
A unified, standardized version of the MIRACL dataset optimized for seamless integration with the Hugging Face ecosystem.
Overview
While the original MIRACL dataset is excellent, loading it through the Hugging Face datasets library often requires running conversion scripts, which can take considerable time before the data becomes usable. To address this, we have pre-converted the entire MIRACL dataset into a format that's immediately ready for use with the Hugging Face datasets library.
MIRACL Unified provides multilingual information retrieval data in a clean, standardized format across 18 languages. This dataset transforms the original MIRACL format into a more accessible structure that's optimized for modern information retrieval research and development.
Key improvements over the original MIRACL dataset:
- 🔄 Unified format across all languages and splits
- 🚀 Direct document indexing with integer IDs for faster retrieval
- 📊 Consistent schema for seamless multi-language processing
- ⚡ Optimized for Hugging Face datasets library
- 🛠️ Ready-to-use for training and evaluation
Supported Languages
The dataset covers 18 languages with comprehensive Wikipedia-based corpora:
- Arabic (ar), Bengali (bn), English (en), Spanish (es)
- Persian (fa), Finnish (fi), French (fr), Hindi (hi)
- Indonesian (id), Japanese (ja), Korean (ko), Russian (ru)
- Swahili (sw), Telugu (te), Thai (th), Chinese (zh)
- German (de), Yoruba (yo)
Dataset Structure
Each language provides two complementary datasets:
Query Datasets ({lang}_queries)
Contains question-answer pairs with relevance annotations:
query_id: Unique identifier for the queryquery: The search question or information needpositives: List of relevant document row indices (integers)negatives: List of non-relevant document row indices (integers)
Available splits: train, dev, testA, testB (varies by language)
Corpus Datasets ({lang}_corpus)
Contains the searchable document collection:
docid: Original MIRACL document identifiertitle: Document titletext: Full document content
Quick Start
from datasets import load_dataset
lang = "ja" # Choose from: en, ar, bn, es, fa, fi, fr, hi, id, ko, ru, sw, te, th, zh, de, yo
# Load queries and corpus
miracl_queries = load_dataset("hotchpotch/miracl-hf-unified", f"{lang}_queries")
miracl_corpus = load_dataset("hotchpotch/miracl-hf-unified", f"{lang}_corpus", split="train")
# Check available splits
print(f"Available query splits: {list(miracl_queries.keys())}") # e.g., ['train', 'dev', 'testA', 'testB']
# Get a query from any split (train, dev, testA, etc.)
query = miracl_queries["train"][0]
print(f"Query: {query['query']}")
# Extract positive and negative documents using row indices
positive_docs = [miracl_corpus[row_id] for row_id in query["positives"]]
negative_docs = [miracl_corpus[row_id] for row_id in query["negatives"]]
print(f"Found {len(positive_docs)} relevant and {len(negative_docs)} non-relevant docs")
Dataset Statistics
| Language | Corpus Size | Train Queries | Dev Queries |
|---|---|---|---|
| Arabic (ar) | 2.1M | 3,495 | 2,896 |
| Bengali (bn) | 297K | 1,631 | 411 |
| English (en) | 32.9M | 2,863 | 799 |
| Spanish (es) | 10.4M | 2,162 | 648 |
| Persian (fa) | 2.2M | 2,107 | 683 |
| Finnish (fi) | 1.9M | 2,897 | 1,271 |
| French (fr) | 14.6M | 1,143 | 343 |
| Hindi (hi) | 506K | 1,169 | 350 |
| Indonesian (id) | 1.4M | 4,071 | 960 |
| Japanese (ja) | 7.0M | 3,477 | 860 |
| Korean (ko) | 1.5M | 868 | 213 |
| Russian (ru) | 9.5M | 4,683 | 1,252 |
| Swahili (sw) | 132K | 1,901 | 482 |
| Telugu (te) | 518K | 3,452 | 646 |
| Thai (th) | 542K | 2,972 | 807 |
| Chinese (zh) | 4.9M | 1,312 | 393 |
| German (de) | 15.9M | - | 305 |
| Yoruba (yo) | 49K | - | 119 |
Note: Some languages only have dev/test splits without training data. Be aware that in MIRACL evaluation practices, the "dev" split is often used as the actual test set for evaluation.
Key Features
🎯 Direct Document Access
Unlike the original MIRACL format, positive and negative documents are referenced by integer row indices that directly index into the corpus, enabling instant document retrieval without complex mapping.
🔄 Consistent Schema
All languages follow the same data structure, making it easy to build multi-lingual systems and conduct cross-lingual experiments.
⚡ Optimized Performance
- Pre-computed document mappings for faster access
- Efficient storage format compatible with Hugging Face datasets
- Memory-efficient loading options for large corpora
🛠️ Research-Ready
Perfect for:
- Dense passage retrieval training and evaluation
- Cross-lingual information retrieval research
- Multi-lingual question answering systems
- Retrieval-augmented generation (RAG) applications
- Zero-shot cross-lingual transfer studies
Citation
If you use this dataset, please cite the original MIRACL paper:
@article{zhang2022miracl,
title={MIRACL: A Multilingual Information Retrieval Across Languages Dataset},
author={Zhang, Xinyu and Thakur, Nandan and Ogundepo, Odunayo and Kamalloo, Ehsan and Alfonso-Hermelo, David and Li, Xiaoguang and Liu, Qun and Rezagholizadeh, Mehdi and Lin, Jimmy},
journal={arXiv preprint arXiv:2210.09984},
year={2022}
}
License
This dataset is released under the Apache 2.0 License, same as the original MIRACL dataset.
Related Resources
- 📄 Original MIRACL Dataset: https://huggingface.co/datasets/miracl/miracl
- 📝 MIRACL Paper: https://arxiv.org/abs/2210.09984
Acknowledgments
This unified dataset is built upon the excellent work of the MIRACL team. We thank the original authors for creating such a valuable resource for the multilingual information retrieval community.
Author
Yuichi Tateno (@hotchpotch)