Dataset Card for NoMIRACL (EMNLP 2024 Findings Track)

Quick Overview

This repository contains the topics, qrels, and top-k (a maximum of 10) annotated passages. The passage collection can be found here on HF: miracl/miracl-corpus.

import datasets

language = 'german'  # or any of the 18 languages (mentioned above in `languages`)
subset = 'relevant'  # or 'non_relevant' (two subsets: relevant & non-relevant)
split = 'test'       # or 'dev' for the development split

# four combinations available: 'dev.relevant', 'dev.non_relevant', 'test.relevant' and 'test.non_relevant'
nomiracl = datasets.load_dataset('miracl/nomiracl', language, split=f'{split}.{subset}', trust_remote_code=True)

What is NoMIRACL?

Retrieval Augmented Generation (RAG) is a powerful approach to incorporating external knowledge into large language models (LLMs) to enhance the accuracy and faithfulness of LLM-generated responses. However, evaluating query-passage relevance across diverse language families has been a challenge, leading to gaps in understanding the model's performance against errors in external retrieved knowledge. To address this, we present NoMIRACL, a completely human-annotated dataset designed for evaluating multilingual LLM relevance across 18 diverse languages.

NoMIRACL evaluates LLM relevance as a binary classification objective, containing two subsets: non-relevant and relevant. The non-relevant subset contains queries with all passages manually judged by an expert assessor as non-relevant, while the relevant subset contains queries with at least one judged relevant passage within the labeled passages. LLM relevance is measured using two key metrics:

• hallucination rate (on the non-relevant subset) measuring model tendency to recognize when none of the passages provided are relevant for a given question (non-answerable).
• error rate (on the relevant subset) measuring model tendency as unable to identify relevant passages when provided for a given question (answerable).

Acknowledgement

This dataset would not have been possible without all the topics are generated by native speakers of each language in conjunction with our multilingual RAG universe work in part 1, MIRACL [TACL '23]. The queries with all non-relevant passages are used to create the non-relevant subset whereas queries with at least a single relevant passage (i.e., MIRACL dev and test splits) are used to create relevant subset.

This repository contains the topics, qrels and top-10 (maximum) annotated documents of NoMIRACL. The whole collection can be found here.

Quickstart

import datasets

language = 'german'  # or any of the 18 languages
subset = 'relevant'  # or 'non_relevant'
split = 'test'       # or 'dev' for development split

# four combinations available: 'dev.relevant', 'dev.non_relevant', 'test.relevant' and 'test.non_relevant'
nomiracl = datasets.load_dataset('miracl/nomiracl', language, split=f'{split}.{subset}', trust_remote_code=True)

Dataset Structure

1. To download the files:

Under folders data/{lang}, the subset of the corpus is saved in .jsonl.gz format, with each line to be:

{"docid": "28742#27", 
"title": "Supercontinent", 
"text": "Oxygen levels of the Archaean Eon were negligible and today they are roughly 21 percent. [ ... ]"}

Under folders data/{lang}/topics, the topics are saved in .tsv format, with each line to be:

qid\tquery

Under folders miracl-v1.0-{lang}/qrels, the qrels are saved in standard TREC format, with each line to be:

qid Q0 docid relevance

2. To access the data using HuggingFace datasets:

import datasets

language = 'german'  # or any of the 18 languages
subset = 'relevant'  # or 'non_relevant'
split = 'test'       # or 'dev' for development split

# four combinations: 'dev.relevant', 'dev.non_relevant', 'test.relevant' and 'test.non_relevant'
nomiracl = datasets.load_dataset('miracl/nomiracl', language, split=f'{split}.{subset}')

# Individual entry in `relevant` or `non_relevant` subset
for data in nomiracl:  # or 'dev', 'testA'
  query_id = data['query_id']
  query = data['query']
  positive_passages = data['positive_passages']
  negative_passages = data['negative_passages']
  
  for entry in positive_passages: # OR 'negative_passages'
    docid = entry['docid']
    title = entry['title']
    text = entry['text']

Dataset Statistics

For NoMIRACL dataset statistics, please refer to our EMNLP 2024 Findings publication.

Paper: https://aclanthology.org/2024.findings-emnlp.730/.

Citation Information

This work was conducted as a collaboration between the University of Waterloo and Huawei Technologies.

@inproceedings{thakur-etal-2024-knowing,
    title = "{``}Knowing When You Don{'}t Know{''}: A Multilingual Relevance Assessment Dataset for Robust Retrieval-Augmented Generation",
    author = "Thakur, Nandan  and
      Bonifacio, Luiz  and
      Zhang, Crystina  and
      Ogundepo, Odunayo  and
      Kamalloo, Ehsan  and
      Alfonso-Hermelo, David  and
      Li, Xiaoguang  and
      Liu, Qun  and
      Chen, Boxing  and
      Rezagholizadeh, Mehdi  and
      Lin, Jimmy",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.730",
    pages = "12508--12526",
    abstract = "Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) output by leveraging external knowledge sources to reduce factual hallucinations. However, prior work lacks a comprehensive evaluation of different language families, making it challenging to evaluate LLM robustness against errors in external retrieved knowledge. To overcome this, we establish **NoMIRACL**, a human-annotated dataset for evaluating LLM robustness in RAG across 18 typologically diverse languages. NoMIRACL includes both a non-relevant and a relevant subset. Queries in the non-relevant subset contain passages judged as non-relevant, whereas queries in the relevant subset include at least a single judged relevant passage. We measure relevance assessment using: (i) *hallucination rate*, measuring model tendency to hallucinate when the answer is not present in passages in the non-relevant subset, and (ii) *error rate*, measuring model inaccuracy to recognize relevant passages in the relevant subset. In our work, we observe that most models struggle to balance the two capacities. Models such as LLAMA-2 and Orca-2 achieve over 88{\%} hallucination rate on the non-relevant subset. Mistral and LLAMA-3 hallucinate less but can achieve up to a 74.9{\%} error rate on the relevant subset. Overall, GPT-4 is observed to provide the best tradeoff on both subsets, highlighting future work necessary to improve LLM robustness. NoMIRACL dataset and evaluation code are available at: https://github.com/project-miracl/nomiracl.",
}