Datasets:
language:
- en
size_categories:
- 10M<n<100M
task_categories:
- question-answering
pretty_name: v
dataset_info:
features:
- name: text
dtype: string
- name: title
dtype: string
splits:
- name: train
num_bytes: 13505907291
num_examples: 21035236
download_size: 7465667343
dataset_size: 13505907291
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
Wikipedia Dump with Gold Documents from Natural Questions
Dataset Summary
This dataset combines the English Wikipedia dump from December 20, 2018, with gold passages from the Natural Questions (NQ) dataset, specifically tailored for open-domain question answering tasks. By integrating gold documents corresponding to each query in the NQ-open version of the dataset, this resource addresses potential mismatches between the Wikipedia dump and the question-answer pairs found in NQ-open. Such mismatches can lead to scenarios where the dump does not contain the required answer. A thorough process of duplicate filtering was applied to ensure the precise identification of the gold document for each query, enhancing the reliability of the dataset for natural language processing tasks.
Therefore, the dataset can be employed as a knowledge base for RAG systems. One critical aspect of dataset preparation involved addressing the constraints posed by Large Language Models (LLMs) regarding input size. LLMs, particularly when processing multiple documents in a single prompt, face limitations on the length of input they can efficiently handle. To accommodate this, gold documents exceeding 512 tokens (tokenized with Llama2) were excluded from the dataset. This decision was guided by the objective of maximizing the number of documents that can be included in the LLM's prompt without compromising on the detail or context provided by each document. As a result, the final dataset encompasses 21,035,236 documents (13.9 GB).
Dataset Sources
Original Wikipedia Dump: The corpus originates from the English Wikipedia dump, where articles are segmented into non-overlapping passages of 100 words. Download link.
Gold Passages: Sourced from the Natural Questions dataset, these passages are integrated to provide a comprehensive resource for question answering. The gold passages are accessible through the following URLs:
The above data comes from the Dense Passage Retrieval (DPR) github repository.
Dataset Structure
An example of a Wikipedia passage is as follows:
{
"text": Home computers were a class of microcomputers entering the market in 1977, and becoming common during the 1980s.
They were marketed to consumers as affordable and accessible computers that, for the first time, were intended for the use of a single nontechnical user.
These computers were a distinct market segment that typically cost much less than business,
scientific or engineering-oriented computers of the time such as the IBM PC, and were generally less powerful in terms of memory and expandability.
However, a home computer often had better graphics and sound than contemporary business computers. Their most common uses were playing
"title": "Home computer"
}