MIRACL Triplets in 51 Languages

Dataset Summary

MIRACL Triplets in 51 Languages is a multilingual retrieval and reranking dataset for training dense retrieval, sentence embedding, reranker, and cross-lingual search models. It contains (query, positive, negative) triplets derived from the English training split of miracl/miracl.

The dataset provides 51 language subsets: the original English data and translations into 50 languages. Each subset contains 2,863 aligned training examples. Rows with the same id represent translations of the same source triplet. The aligned (query, positive, negative) structure supports multilingual retrieval, multilingual reranking, contrastive learning, and hard-negative training.

• Format: Parquet
• Subsets: 51 language codes, including en, ko, ja, zh, de, and fr
• Split: train
• Rows per subset: 2,863
• Total rows across subsets: 146,013

Data Structure

Data Instances

Example from the English subset:

{
  "id": 232,
  "query": "What is Nigers anthem?",
  "positive": "La Nigérienne\n\"La Nigérienne\" is the national anthem of Niger. The lyrics are by Maurice Albert Thiriet. Robert Jacquet and Nicolas Abel François Frionnet wrote the music. It was adopted as Niger's anthem in 1961.",
  "negative": "Nigeria, We Hail Thee\nThe second national anthem, \"Arise, O Compatriots,\" replaced \"Nigeria, We Hail Thee\" in 1978."
}

How to Use

from datasets import load_dataset

# Load one language subset. Replace "ko" with any supported language code.
dataset = load_dataset("nlpai-lab/miracl-multilingual-triplets", "ko", split="train")

print(dataset.features)
print(dataset[0])

The default subset is English (en).

Source and Translation

• Source dataset: miracl/miracl
• Source configuration: en
• Source split: train
• Translation model: google/translategemma-27b-it

The model's original translation prompt template was used without modification or additional instructions. The query, positive, and negative fields were translated independently while preserving their triplet relationship and cross-language id alignment.

Languages

License

The source MIRACL dataset is released under the Apache License 2.0. This derived dataset is also released under Apache-2.0.