Datasets are taken from Facebook's CRAG: Comprehensive RAG Benchmark, see their arXiv paper for details about the dataset construction.

CRAG Sampler

We have added a simple Python tool for performing stratified sampling on CRAG data.

Installation

Local Development Install (Recommended)

git clone https://huggingface.co/Quivr/CRAG.git
cd CRAG
pip install -r requirements.txt  # Install dependencies
pip install -e .  # Install package in development mode

Quick Start

Running the example

python -m examples.basic_sampling

CRAG dataset

CRAG (Comprehensive RAG Benchmark) is a rich and comprehensive factual question answering benchmark designed to advance research in RAG. The public version of the dataset includes:

• 2706 Question-Answer pairs
• 5 domains: Finance, Sports, Music, Movie, and Open domain
• 8 types of questions (see image below): simple, simple with condition, set, comparison, aggregation, multi-hop, post-processing heavy, and false premise

The datasets crag_task_1_and_2_dev_v4_subsample_*.json.bz2 have been created from the dataset crag_task_1_and_2_dev_v4.jsonl.bz2 available on CRAG's GitHub repository. For an easier handling and download of the dataset, we have used our CRAG sampler to split the 2706 rows of the original file in 5 subsamples, following the procedure below:

1. We have created a new label answer_type, classifying the answers in 3 categories:
   • invalid for any answer == "invalid question"
   • no_answer for any answer == "i don't know"
   • valid for any other answer
2. We have considered the labels answer_type, domain, question_type and static_or_dynamic and performed stratified sampling, splitting the datasets in 5 subsamples. Each subsample has thus the same statistical properties of the full dataset.

We report below the data schema as provided in CRAG's GitHub repository.

Data Schema

Field Name	Type	Description
interaction_id	string	A unique identifier for each example.
query_time	string	Date and time when the query and the web search occurred.
domain	string	Domain label for the query. Possible values: "finance", "music", "movie", "sports", "open". "Open" includes any factual queries not among the previous four domains.
question_type	string	Type label about the query. Possible values include: "simple", "simple_w_condition", "comparison", "aggregation", "set", "false_premise", "post-processing", "multi-hop".
static_or_dynamic	string	Indicates whether the answer to a question changes and the expected rate of change. Possible values: "static", "slow-changing", "fast-changing", and "real-time".
query	string	The question for RAG to answer.
answer	string	The gold standard answer to the question.
alt_ans	list	Other valid gold standard answers to the question.
split	integer	Data split indicator, where 0 is for validation and 1 is for the public test.
search_results	list of JSON	Contains up to k HTML pages for each query (k=5 for Task #1 and k=50 for Task #3), including page name, URL, snippet, full HTML, and last modified time.

Search Results Detail

Key	Type	Description
page_name	string	The name of the webpage.
page_url	string	The URL of the webpage.
page_snippet	string	A short paragraph describing the major content of the page.
page_result	string	The full HTML of the webpage.
page_last_modified	string	The time when the page was last modified.