inria-soda/TILE-benchmark

TILE Benchmark: 100 tables with strings

The TILE benchmark gathers 100 tables that mix string and numerical columns, for the study of tabular learning with strings. Each table is a supervised learning problem: 65 regressions, 17 binary classifications and 18 multiclass classifications. They come from three public sources:

Source	Tables	Origin of the tables
CARTE	51	the CARTE benchmark, collected from public sources
TextTabBench	17	Kaggle datasets with long text fields
WikiDBs	32	tables built from Wikidata

The tables range from 1,062 to 297,934 rows (median 9,615), and from 3 to 102 columns besides the target: 2.0 million rows and 206 MB in total.

These tables are the benchmark of the paper TILE: Bridging LLMs and Table Foundation Models for Tabular Learning with Strings. The code to run TILE on them is at https://github.com/soda-inria/tile.

Structure

Each table is a folder named after it, with two files:

• data.parquet: the table, the target included;
• config.json: the name of the target column (target_name), the task (regression, b-classification or m-classification) and the source.

The names of the tables start with their source: carte_, ttb_ (TextTabBench) or wdbs_ (WikiDBs). This is the layout of the STRABLE benchmark, so that the same code reads both.

Download every table:

hf download inria-soda/TILE-benchmark --repo-type dataset --local-dir data/tile

or load one with the datasets library:

from datasets import load_dataset

table = load_dataset("inria-soda/TILE-benchmark", "carte_wina_pl", split="train").to_pandas()

Preprocessing

CARTE. The preprocessed single tables released with the CARTE benchmark.

TextTabBench. The tables were downloaded from Kaggle, then:

• the columns with more than half of their values missing (up to 63% for a few tables), and the columns with a single value, were dropped;
• duplicate rows, and rows with a missing target, were dropped;
• index columns, and columns that leak the target, were removed by hand, along with table-specific cleaning;
• the four largest tables were subsampled at random before cleaning: to 12,000 rows for ttb_calif_houses and ttb_mercari, and to 24,000 rows for ttb_customer_complaints and ttb_kickstarter;
• ttb_spotify keeps the 10 most common music genres;
• the classification targets were label-encoded.

WikiDBs. One table was taken from each of 32 databases, with a column as the target. Following TextTabBench, the columns with more than half of their values missing, the columns with a single value, and duplicate rows were dropped. For classification, the classes with fewer than 105 rows were removed, and at most the 10 most frequent classes kept. For regression, the rows with a non-finite target were removed.

All sources. The classification tables were balanced: every class was subsampled at random to the size of the smallest. A table was kept only if more than 1,050 rows remained.

The string cells are left as they are: the TILE code replaces line breaks and tabs within cells by spaces when it loads a table.

Evaluation protocol

The paper trains on 1,024, 10,000 or 50,000 rows drawn at random (stratified for classification), and tests on up to 1,024 of the remaining rows, over 5 random seeds. It reports R² for regression and ROC AUC for classification.

Licenses

The tables keep the license of their source. The WikiDBs tables are under CC BY 4.0. The license of each TextTabBench table is the one declared on its Kaggle page, listed below. For the CARTE tables, see the CARTE benchmark and the original source of each table. The preprocessing, the configuration files and this card are released under CC BY 4.0.

Citation

If you use this benchmark, please cite the TILE paper, and the source of the tables you use (CARTE, TextTabBench and WikiDBs):

@article{lefebvre2026tile,
  title={TILE: Bridging LLMs and Table Foundation Models for Tabular Learning with Strings},
  author={Lefebvre, F{\'e}lix and Holzm{\"u}ller, David and Blayer, Gioia and Kim, Myung Jun and Morvan, Marine Le and Varoquaux, Ga{\"e}l},
  journal={arXiv preprint arXiv:2610.17065},
  year={2026}
}

@inproceedings{kim2024carte,
  title={{CARTE}: Pretraining and Transfer for Tabular Learning},
  author={Kim, Myung Jun and Grinsztajn, Leo and Varoquaux, Gael},
  booktitle={Proceedings of the 41st International Conference on Machine Learning},
  series={Proceedings of Machine Learning Research},
  volume={235},
  pages={23843--23866},
  publisher={PMLR},
  year={2024}
}

@inproceedings{mraz2025towards,
  title={Towards Benchmarking Foundation Models for Tabular Data With Text},
  author={Mr{\'a}z, Martin and Das, Breenda and Gupta, Anshul and Purucker, Lennart and Hutter, Frank},
  booktitle={1st ICML Workshop on Foundation Models for Structured Data},
  year={2025},
  note={arXiv:2507.07829}
}

@article{vogel2024wikidbs,
  title={{WikiDBs}: A large-scale corpus of relational databases from {Wikidata}},
  author={Vogel, Liane and Bodensohn, Jan-Micha and Binnig, Carsten},
  journal={Advances in Neural Information Processing Systems},
  volume={37},
  pages={41186--41201},
  year={2024}
}

The tables

Table	Name	Task	Target	Rows	Columns	Origin	License
carte_anime_planet	Anime Planet	regression	Rating_Score	14,391	10	CARTE	see CARTE
carte_babies_r_us	Babies R Us	regression	price	5,085	4	CARTE	see CARTE
carte_beer_ratings	Beer Ratings	regression	review_overall	3,197	19	CARTE	see CARTE
carte_bikedekho	BikeDekho	regression	price	4,786	7	CARTE	see CARTE
carte_bikewale	BikeWale	regression	price	8,992	7	CARTE	see CARTE
carte_buy_buy_baby	Buy Buy Baby	regression	price	10,718	4	CARTE	see CARTE
carte_cardekho	CarDekho	regression	price	37,813	16	CARTE	see CARTE
carte_chocolate_bar_ratings	Chocolate Bar Ratings	binary classification	Rating	2,218	8	CARTE	see CARTE
carte_clear_corpus	CLEAR Corpus	regression	BT_Easiness	4,724	26	CARTE	see CARTE
carte_coffee_ratings	Coffee Ratings	binary classification	rating	1,670	8	CARTE	see CARTE
carte_company_employees	Company Employees	regression	current_employee_estimate	10,941	7	CARTE	see CARTE
carte_employee_remuneration	Employee Remuneration	regression	Remuneration	35,396	4	CARTE	see CARTE
carte_employee_salaries	Employee Salaries	regression	current_annual_salary	9,211	8	CARTE	see CARTE
carte_fifa22_players	FIFA 22 Players	regression	wage_eur	18,085	18	CARTE	see CARTE
carte_filmtv_movies	FilmTV Movies	regression	public_vote	41,205	9	CARTE	see CARTE
carte_journal_jcr	Journal JCR	regression	2021_JIF	9,615	9	CARTE	see CARTE
carte_journal_sjr	Journal SJR	regression	H_index	27,931	9	CARTE	see CARTE
carte_jp_anime	Japanese Anime	regression	Score	15,535	16	CARTE	see CARTE
carte_k_drama	K-Drama	regression	score	1,239	12	CARTE	see CARTE
carte_michelin	Michelin	binary classification	Award	6,774	7	CARTE	see CARTE
carte_mlds_salaries	ML/DS Salaries	regression	salary_in_usd	10,456	8	CARTE	see CARTE
carte_movies	Movies	regression	revenue	7,224	13	CARTE	see CARTE
carte_museums	Museums	regression	Revenue	11,467	16	CARTE	see CARTE
carte_mydramalist	MyDramaList	regression	rating	3,400	13	CARTE	see CARTE
carte_nba_draft	NBA Draft	binary classification	value_over_replacement	1,550	6	CARTE	see CARTE
carte_prescription_drugs	Prescription Drugs	regression	WAC_at_Introduction	1,714	7	CARTE	see CARTE
carte_ramen_ratings	Ramen Ratings	binary classification	Stars	3,726	4	CARTE	see CARTE
carte_roger_ebert	Roger Ebert	binary classification	critic_rating	2,668	7	CARTE	see CARTE
carte_rotten_tomatoes	Rotten Tomatoes	regression	Rating_Value	7,158	13	CARTE	see CARTE
carte_spotify	Spotify	binary classification	popularity	41,096	18	CARTE	see CARTE
carte_us_accidents_counts	US Accidents Counts	regression	Counts	22,623	6	CARTE	see CARTE
carte_us_accidents_severity	US Accidents Severity	binary classification	Severity	20,930	10	CARTE	see CARTE
carte_us_presidential	US Presidential	regression	target	19,857	6	CARTE	see CARTE
carte_used_cars_24	Used Cars 24	regression	Price	5,918	8	CARTE	see CARTE
carte_used_cars_benz_italy	Used Cars Benz Italy	regression	price	16,391	7	CARTE	see CARTE
carte_used_cars_dot_com	UsedCars.Com	regression	price	4,009	10	CARTE	see CARTE
carte_used_cars_pakistan	Used Cars Pakistan	regression	Price	72,655	7	CARTE	see CARTE
carte_used_cars_saudi_arabia	Used Cars Saudi Arabia	regression	Price	5,507	10	CARTE	see CARTE
carte_videogame_sales	Videogame Sales	regression	Global_Sales	16,410	5	CARTE	see CARTE
carte_whisky	Whisky	binary classification	Meta_Critic	1,788	6	CARTE	see CARTE
carte_wikiliq_beer	WikiLIQ Beer	regression	Price	13,461	9	CARTE	see CARTE
carte_wikiliq_spirit	WikiLIQ Spirit	regression	Price	12,275	7	CARTE	see CARTE
carte_wina_pl	Wina Poland	regression	price	2,247	15	CARTE	see CARTE
carte_wine_dot_com_prices	Wine.com Prices	regression	Prices	15,254	9	CARTE	see CARTE
carte_wine_dot_com_ratings	Wine.com Ratings	regression	Ratings	4,095	9	CARTE	see CARTE
carte_wine_enthusiasts_prices	WineEnthusiasts Prices	regression	price	120,975	9	CARTE	see CARTE
carte_wine_enthusiasts_ratings	WineEnthusiasts Ratings	regression	points	129,971	9	CARTE	see CARTE
carte_wine_vivino_price	Vivino Prices	regression	Price	13,834	7	CARTE	see CARTE
carte_wine_vivino_rating	Vivino Ratings	regression	Rating	13,834	8	CARTE	see CARTE
carte_yelp	Yelp	binary classification	stars	60,088	12	CARTE	see CARTE
carte_zomato	Zomato	binary classification	rating	60,302	8	CARTE	see CARTE
ttb_airbnb	Airbnb	regression	price	3,818	55	Kaggle	CC0-1.0
ttb_beer	Beer	regression	review_overall	2,914	20	Kaggle	CC BY 4.0
ttb_calif_houses	California Houses	regression	Total interior livable area	11,349	29	Kaggle	Kaggle competition rules
ttb_covid_trials	COVID Trials	regression	days_to_complete	1,165	15	Kaggle	DbCL-1.0
ttb_customer_complaints	Customer Complaints	multiclass classification	Company response to consumer	1,384	10	Kaggle	U.S. Government Works
ttb_diabetes	Diabetes	binary classification	diabetes	17,000	16	Kaggle	CC0-1.0
ttb_insurance_complaints	Insurance Complaints	regression	Recovery	37,484	10	Kaggle	CC0-1.0
ttb_it_salary	IT Salary	regression	Yearly brutto salary (without bonus and stocks) in EUR	1,253	18	Kaggle	CC0-1.0
ttb_job_frauds	Job Frauds	binary classification	fraudulent	1,732	14	Kaggle	CC0-1.0
ttb_kickstarter	Kickstarter	binary classification	status	18,720	17	Kaggle	unknown
ttb_lending_club	Lending Club	binary classification	loan_status	11,254	27	Kaggle	CC0-1.0
ttb_mercari	Mercari	regression	price	12,000	6	Kaggle	CC0-1.0
ttb_osha_accidents	OSHA Accidents	binary classification	Task Assigned	3,598	18	Kaggle	unknown
ttb_sf_permits	San Francisco Permits	regression	time_to_approve	183,794	28	Kaggle	DbCL-1.0
ttb_spotify	Spotify	multiclass classification	track_genre	10,000	18	Kaggle	ODbL-1.0
ttb_stack_overflow	Stack Overflow	regression	Yearly_Salary_USD	19,427	102	Kaggle	Apache-2.0
ttb_wine	Wine	regression	Price	1,281	15	Kaggle	CC0-1.0
wdbs_03977_Eclipsing_Binary_Star_Instances	Eclipsing Binary Stars	regression	Apparent_Magnitude	297,934	8	WikiDBs	CC BY 4.0
wdbs_07310_DecommissionedTransportStations	Transport Stations	multiclass classification	Country	4,640	9	WikiDBs	CC BY 4.0
wdbs_07900_ArtistCopyrightRepresentation	Artists Copyright	multiclass classification	ArtistOccupation	2,000	10	WikiDBs	CC BY 4.0
wdbs_09510_CreativeCommonsAuthors	CC Authors	binary classification	Gender	16,224	8	WikiDBs	CC BY 4.0
wdbs_14012_ResearchArticleCitations	Research Article Citations	regression	PublicationDate	4,115	9	WikiDBs	CC BY 4.0
wdbs_14976_DrawingsCatalog	Drawings Catalog	regression	ArtworkHeightCm	63,130	9	WikiDBs	CC BY 4.0
wdbs_15542_FORWARD_PLAYERS	Forward Players	multiclass classification	SPORTS_TEAM	1,400	10	WikiDBs	CC BY 4.0
wdbs_19664_MunicipalDistrictCapitals	Municipal District Capitals	regression	PopulationCount	1,658	8	WikiDBs	CC BY 4.0
wdbs_28324_ukrainian_village_instances	Ukrainian Villages	regression	elevation_meters	21,355	6	WikiDBs	CC BY 4.0
wdbs_30417_ArtworksCatalog	Artworks Catalog	multiclass classification	ArtworkType	1,210	9	WikiDBs	CC BY 4.0
wdbs_36100_MagicNarrativeMotifs	Magic Narratives	multiclass classification	CulturalOrigin	1,062	4	WikiDBs	CC BY 4.0
wdbs_42562_Geographer_Profiles	Geographers	multiclass classification	Languages	1,130	7	WikiDBs	CC BY 4.0
wdbs_46159_DissolvedMunicipalityRecords	Dissolved Municipalities	regression	DissolutionDate	13,462	8	WikiDBs	CC BY 4.0
wdbs_53353_research_articles	Research Articles	regression	publication_date	6,962	8	WikiDBs	CC BY 4.0
wdbs_56474_Sculpture_Instances	Sculptures	multiclass classification	Material_Used	3,720	6	WikiDBs	CC BY 4.0
wdbs_62826_HISTORICAL_FIGURES	Historical Figures	regression	BIRTH_DATE	11,260	11	WikiDBs	CC BY 4.0
wdbs_63797_SpringLocations	Spring Locations	multiclass classification	CountryName	5,930	3	WikiDBs	CC BY 4.0
wdbs_64477_NobleIndividuals	Noble Individuals	multiclass classification	Role	1,400	9	WikiDBs	CC BY 4.0
wdbs_65102_defender_profiles	Defenders	multiclass classification	citizenship_country	18,610	8	WikiDBs	CC BY 4.0
wdbs_66610_geopolitical_regions	Geopolitical Regions	regression	land_area	1,114	9	WikiDBs	CC BY 4.0
wdbs_70780_StateSchoolDetails	State Schools	multiclass classification	Country	2,800	4	WikiDBs	CC BY 4.0
wdbs_70942_NotableTreesInformation	Notable Trees	multiclass classification	TreeSpecies	1,408	6	WikiDBs	CC BY 4.0
wdbs_82939_Territorial_Entities	Territorial Entities	regression	Population_Count	36,717	11	WikiDBs	CC BY 4.0
wdbs_88197_artworks_inventory	Artworks Inventory	regression	artwork_width_cm	10,635	6	WikiDBs	CC BY 4.0
wdbs_89039_Business_Entity_Locations	Business Locations	regression	Population_Count	16,821	8	WikiDBs	CC BY 4.0
wdbs_89439_WwiPersonnelProfiles	WWI Personnel	regression	BirthDate	30,675	11	WikiDBs	CC BY 4.0
wdbs_90741_MUSEUM_DETAILS	Museums	multiclass classification	COUNTRY	9,550	6	WikiDBs	CC BY 4.0
wdbs_90930_RegisteredShips	Registered Ships	regression	GrossTonnage	4,644	9	WikiDBs	CC BY 4.0
wdbs_92415_island_details	Islands	multiclass classification	country_name	19,650	4	WikiDBs	CC BY 4.0
wdbs_94062_POET_PROFILES	Poets	regression	DEATH_DATE	60,240	10	WikiDBs	CC BY 4.0
wdbs_97229_PhilosopherProfiles	Philosophers	multiclass classification	Languages	4,230	7	WikiDBs	CC BY 4.0
wdbs_97297_MusicAlbumsPublishedInUs	US Music Albums	multiclass classification	MusicGenre	3,270	11	WikiDBs	CC BY 4.0