You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

AfriXIR

Dataset Summary

AfriXIR is a multilingual passage retrieval dataset for training and evaluating retrieval models in eight African languages. The dataset contains 112,864 training examples and 12,546 test examples. Each example is a query, its positive passage, and a list of hard negative passages. Inside one row, the query, the positive passage, and the negative passages are in that row's language.

Languages

Language ISO Train Test Share
Amharic am 61,469 6,832 54.5%
Swahili sw 14,117 1,569 12.5%
Hausa ha 9,972 1,108 8.8%
Yoruba yo 8,956 996 7.9%
Oromo om 8,259 918 7.3%
Somali so 8,000 889 7.1%
Shona sn 1,656 185 1.5%
Tigrinya ti 435 49 0.4%

Total: 112,864 train + 12,546 test = 125,410 rows

from datasets import load_dataset

afrixir = load_dataset("nebhailema/AfriXIR")

Data Collection

Query-passage pairs come from these sources:

  1. MasakhaNEWS for Hausa, Oromo, Shona, Somali, Swahili, and Yoruba. The headline is the query and the article is the passage. masakhane/masakhanews

  2. XL-Sum for Hausa, Oromo, Somali, Swahili, and Yoruba. The summary is the query and the article is the passage. csebuetnlp/xlsum

  3. KenSwQuAD for Swahili. The question is the query and the story is the passage. Kencorpus/KenSwQuAD

  4. Mr. TyDi for Swahili. The question is the query and the Wikipedia passage is the positive passage. castorini/mr-tydi

  5. TiQuAD for Tigrinya. The question is the query and the news paragraph is the passage. fgaim/tiquad

  6. Amharic Passage Retrieval Dataset V2 with negatives for Amharic. Those rows are used as published, including their own negatives. The underlying sources are Amharic news, Amharic XL-Sum, Amharic Wikipedia, and AmharicQA. rasyosef/Amharic-Passage-Retrieval-Dataset-V2-With-Negatives

Shona uses MasakhaNEWS only. Tigrinya uses TiQuAD only.

Hard negatives

For every language except Amharic, passages were embedded with sentence-transformers/paraphrase-multilingual-mpnet-base-v2. The negatives are the passages at similarity ranks 20 through 50. Each negative keeps its cosine similarity score. Amharic keeps the 16 negatives shipped with its source dataset.

Split

The built languages use a 90/10 split by query, with seed 42. Amharic keeps the train split and the test split from its source. The language datasets are then combined.

License

The Amharic rows and the XL-Sum rows are CC-BY-NC-SA-4.0, so the dataset card uses that license. MasakhaNEWS rows are AFL-3.0, KenSwQuAD rows are CC-BY-4.0, TiQuAD rows are CC-BY-SA-4.0, and Mr. TyDi rows are Apache-2.0.

Downloads last month
93

Models trained or fine-tuned on nebhailema/AfriXIR

Collection including nebhailema/AfriXIR