💾🏋️💾 DataBench 💾🏋️💾

This repository contains the original 65 datasets used for the paper Question Answering over Tabular Data with DataBench: A Large-Scale Empirical Evaluation of LLMs which appeared in LREC-COLING 2024.

Large Language Models (LLMs) are showing emerging abilities, and one of the latest recognized ones is tabular reasoning in question answering on tabular data. Although there are some available datasets to assess question answering systems on tabular data, they are not large and diverse enough to evaluate this new ability of LLMs. To this end, we provide a corpus of 65 real world datasets, with 3,269,975 and 1615 columns in total, and 1300 questions to evaluate your models for the task of QA over Tabular Data.

Usage

from datasets import load_dataset

# Load all QA pairs
all_qa = load_dataset("cardiffnlp/databench", name="qa", split="full")

# Load SemEval 2025 task 8 Question-Answer splits
semeval_train_qa = load_dataset("cardiffnlp/databench", name="semeval", split="train")
semeval_dev_qa = load_dataset("cardiffnlp/databench", name="semeval", split="dev")


# "001_Forbes", the id of the dataset where information to answer the Question is located
all_qa['dataset'][0] 

# This id can be used load a specific Question-Answer pair collection from the splits
forbes_qa = load_dataset("cardiffnlp/databench", name="qa", split=all_qa['dataset'][0] )

# you can load a specific dataset containg the "answer" for a QA pair using the 
forbes_full = load_dataset("cardiffnlp/databench", name=all_qa['dataset'][0] , split="full")

# or to load the databench lite equivalent dataset, to answer the "sample_answer"
forbes_sample = load_dataset("cardiffnlp/databench", name=all_qa['dataset'][0] , split="lite")

📚 Datasets

By clicking on each name in the table below, you will be able to explore each dataset.

	Name	Rows	Cols	Domain	Source (Reference)
1	Forbes	2668	17	Business	Forbes
2	Titanic	887	8	Travel and Locations	Kaggle
3	Love	373	35	Social Networks and Surveys	Graphext
4	Taxi	100000	20	Travel and Locations	Kaggle
5	NYC Calls	100000	46	Business	City of New York
6	London Airbnbs	75241	74	Travel and Locations	Kaggle
7	Fifa	14620	59	Sports and Entertainment	Kaggle
8	Tornados	67558	14	Health	Kaggle
9	Central Park	56245	6	Travel and Locations	Kaggle
10	ECommerce Reviews	23486	10	Business	Kaggle
11	SF Police	713107	35	Social Networks and Surveys	US Gov
12	Heart Failure	918	12	Health	Kaggle
13	Roller Coasters	1087	56	Sports and Entertainment	Kaggle
14	Madrid Airbnbs	20776	75	Travel and Locations	Inside Airbnb
15	Food Names	906	4	Business	Data World
16	Holiday Package Sales	4888	20	Travel and Locations	Kaggle
17	Hacker News	9429	20	Social Networks and Surveys	Kaggle
18	Staff Satisfaction	14999	11	Business	Kaggle
19	Aircraft Accidents	23519	23	Health	Kaggle
20	Real Estate Madrid	26026	59	Business	Idealista
21	Telco Customer Churn	7043	21	Business	Kaggle
22	Airbnbs Listings NY	37012	33	Travel and Locations	Kaggle
23	Climate in Madrid	36858	26	Travel and Locations	AEMET
24	Salary Survey Spain 2018	216726	29	Business	INE
25	Data Driven SEO	62	5	Business	Graphext
26	Predicting Wine Quality	1599	12	Business	Kaggle
27	Supermarket Sales	1000	17	Business	Kaggle
28	Predict Diabetes	768	9	Health	Kaggle
29	NYTimes World In 2021	52588	5	Travel and Locations	New York Times
30	Professionals Kaggle Survey	19169	64	Business	Kaggle
31	Trustpilot Reviews	8020	6	Business	TrustPilot
32	Delicatessen Customers	2240	29	Business	Kaggle
33	Employee Attrition	14999	11	Business	Kaggle(modified)
34	World Happiness Report 2020	153	20	Social Networks and Surveys	World Happiness
35	Billboard Lyrics	5100	6	Sports and Entertainment	Brown University
36	US Migrations 2012-2016	288300	9	Social Networks and Surveys	US Census
37	Ted Talks	4005	19	Social Networks and Surveys	Kaggle
38	Stroke Likelihood	5110	12	Health	Kaggle
39	Happy Moments	100535	11	Social Networks and Surveys	Kaggle
40	Speed Dating	8378	123	Social Networks and Surveys	Kaggle
41	Airline Mentions X (former Twitter)	14640	15	Social Networks and Surveys	X (former Twitter)
42	Predict Student Performance	395	33	Business	Kaggle
43	Loan Defaults	83656	20	Business	SBA
44	IMDb Movies	85855	22	Sports and Entertainment	Kaggle
45	Spotify Song Popularity	21000	19	Sports and Entertainment	Spotify
46	120 Years Olympics	271116	15	Sports and Entertainment	Kaggle
47	Bank Customer Churn	7088	15	Business	Kaggle
48	Data Science Salary Data	742	28	Business	Kaggle
49	Boris Johnson UK PM Tweets	3220	34	Social Networks and Surveys	X (former Twitter)
50	ING 2019 X Mentions	7244	22	Social Networks and Surveys	X (former Twitter)
51	Pokemon Features	1072	13	Business	Kaggle
52	Professional Map	1227	12	Business	Kern et al, PNAS'20
53	Google Patents	9999	20	Business	BigQuery
54	Joe Biden Tweets	491	34	Social Networks and Surveys	X (former Twitter)
55	German Loans	1000	18	Business	Kaggle
56	Emoji Diet	58	35	Health	Kaggle
57	Spain Survey 2015	20000	45	Social Networks and Surveys	CIS
58	US Polls 2020	3523	52	Social Networks and Surveys	Brandwatch
59	Second Hand Cars	50000	21	Business	DataMarket
60	Bakery Purchases	20507	5	Business	Kaggle
61	Disneyland Customer Reviews	42656	6	Travel and Locations	Kaggle
62	Trump Tweets	15039	20	Social Networks and Surveys	X (former Twitter)
63	Influencers	1039	14	Social Networks and Surveys	X (former Twitter)
64	Clustering Zoo Animals	101	18	Health	Kaggle
65	RFM Analysis	541909	8	Business	UCI ML

🏗️ Folder structure

Each folder represents one dataset. You will find the following files within:

• all.parquet: the processed data, with each column tagged with our typing system, in parquet.
• qa.parquet: contains the human-made set of questions, tagged by type and columns used, for the dataset (sample_answer indicates the answers for DataBench lite)
• sample.parquet: sample containing 20 rows of the original dataset (DataBench lite)
• info.yml: additional information about the dataset

🗂️ Column typing system

In an effort to map the stage for later analysis, we have categorized the columns by type. This information allows us to segment different kinds of data so that we can subsequently analyze the model's behavior on each column type separately. All parquet files have been casted to their smallest viable data type using the open source Lector reader.

What this means is that in the data types we have more granular information that allows us to know if the column contains NaNs or not (following panda’s convention of Int vs int), as well as whether small numerical values contain negatives (Uint vs int) and their range. We also have dates with potential timezone information (although for now they’re all UTC), as well as information about categories’ cardinality coming from the arrow types.

In the table below you can see all the data types assigned to each column, as well as the number of columns for each type. The most common data types are numbers and categories with 1336 columns of the total of 1615 included in DataBench. These are followed by some other more rare types as urls, booleans, dates or lists of elements.

Type	Columns	Example
number	788	55
category	548	apple
date	50	1970-01-01
text	46	A red fox ran...
url	31	google.com
boolean	18	True
list[number]	14	[1,2,3]
list[category]	112	[apple, orange, banana]
list[url]	8	[google.com, apple.com]

🔗 Reference

You can download the paper here.

If you use this resource, please use the following reference:

@inproceedings{oses-etal-2024-databench,
    title = "Question Answering over Tabular Data with DataBench: A Large-Scale Empirical Evaluation of LLMs",
    author = "Jorge Osés Grijalba and Luis Alfonso Ureña-López and
    Eugenio Martínez Cámara and Jose Camacho-Collados",
    booktitle = "Proceedings of LREC-COLING 2024",
    year = "2024",
    address = "Turin, Italy"
}