mFollowIR-parquet

This is a parquet version of the mFollowIR dataset that can be loaded directly with load_dataset(). The original dataset can be found at jhu-clsp/mFollowIR.

Dataset Structure

The dataset contains the following configurations for each language (fas, rus, zho):

Configurations

• qrels_og_[lang]: Original relevance judgments (test split)
• qrels_changed_[lang]: Modified relevance judgments (test split)
• corpus_[lang]: Document collection
• queries_[lang]: Query set with instructions
• top_ranked_[lang]: Top ranked documents

Usage

from datasets import load_dataset

# Load a specific configuration
dataset = load_dataset("jhu-clsp/mFollowIR-parquet", "queries_fas")  # or any other config

# Load multiple configurations
dataset = load_dataset("jhu-clsp/mFollowIR-parquet", 
                      ["queries_fas", "corpus_fas"])

Citation

@article{weller2024mfollowir,
  title="mFollowIR: a Multilingual Benchmark for Instruction Following in Information Retrieval",
  author="Weller, Orion and Chang, Benjamin and Yang, Eugene and Yarmohammadi, Mahsa and Barham, Sam and MacAvaney, Sean and Cohan, Arman and Soldaini, Luca and Van Durme, Benjamin and Lawrie, Dawn",
  journal="arXiv preprint TODO",
  year="2024"
}