RubikBench

Version: v0.9.2 (2026-02-13)

Database Homepage: RubikBench

RubikBench is an enterprise-scale financial database designed for realistic Natural Language to SQL (NL2SQL) research and evaluation.

The RubikBench database contains the financial data of APEX, a hypothetical international automobile manufacturing and sales company. As a financial database, it is designed to support various analytical queries related to the company's operations, sales, and financial performance. This (imaginary) company operates mainly in China, the United States, and Europe. Therefore the database is bilingual, with both English and Chinese values, and uses three currencies: CNY, USD, and EUR.

While the data values are synthesized, the schemas and structural patterns are closely modeled after actual enterprise financial databases, ensuring practical relevance for NL2SQL system development and evaluation. The database is specifically designed to reflect the complexities encountered in real-world enterprise environments, including wide table schemas, domain-specific knowledge, diverse metrics and data calibers, etc.

Database Statistics

2.1 Overview

• Total Size: ~ 38 GB (DuckDB) / ~ 11 GB (parquet)
• Total Tables: 20 tables
• Total Records: ~ 901.53 million rows
• Time Coverage: Monthly from January 2020 to December 2025 (72 months)
• Languages: Bilingual (English and Chinese)
• Database Format: DuckDB Native (recommended), parquet (for huggingface)
  • Formats like sqlite is not supported as it does not compress the data. The full, uncompressed size of data can be over 500GB.

2.2 Table Statistics

Table Name	Rows	Columns	Period	Region	Customer	Dealer	Product	Contract	Project	Revenue	Expense
INCOME_ALL	12.41M	61	72	✓	✓		✓			✓
INCOME_ACCESSORY	3.37M	61	72	✓	✓		✓			✓
INCOME_CYBVEHSYS	0.99M	61	72	✓	✓		✓			✓
INCOME_GALMOTGRO	1.72M	61	72	✓	✓		✓			✓
INCOME_NOVDYN	1.42M	61	72	✓	✓		✓			✓
INCOME_SERVICE	1.48M	61	72	✓	✓		✓			✓
INCOME_SHIMOTCOR	1.13M	61	72	✓	✓		✓			✓
INCOME_TYRAUTGRO	1.04M	61	72	✓	✓		✓			✓
INCOME_WEYMOTCOR	1.25M	61	72	✓	✓		✓			✓
BUDGET_AND_FORECAST_ALL	96.58M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_ACCESSORY	30.18M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_CYBVEHSYS	8.50M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_GALMOTGRO	11.23M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_NOVDYN	10.46M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_SERVICE	9.48M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_SHIMOTCOR	8.86M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_TYRAUTGRO	8.51M	37	72	✓			✓			✓	✓
BUDGET_AND_FORECAST_WEYMOTCOR	9.35M	37	72	✓			✓			✓	✓
PROFIT_AND_LOSS	161.74M	67	72	✓	✓	✓	✓			✓	✓
SALES_LEDGER	521.81M	55	72	✓	✓	✓	✓	✓	✓	✓	✓

2.3 Entity Statistics

Dimension	Hierarchy Levels	Total Entities
Period	1	72 months (202001-202512)
Region	6	2 overseas/domestic, 11 sales regions, 22 countries, 37 national areas, 46 provinces, 47 cities
Product	4	3 (Lv0), 19 (Lv1), 93 (Lv2), 353 (Lv3)
Customer	4	4 (Lv1), 9 (Lv2), 14 (Lv3), 22 (Lv4)
Dealer	3	3 (Lv1), 5 (Lv2), 9 (Lv3) including 'Direct Sales'
Report Item	4	2 (Lv0), 6 (Lv1), 10 (Lv2), 22 (Lv3)
Contract/Project	2	40,756 distinct contracts, 46,895 distinct projects
Caliber	1	2 calibers (A, B)
Currency	1	3 currencies (CNY, EUR, USD)

2.4 Database Explanation

RubikBench is a database containing the financial data of APEX, an (imaginary) international automobile manufacturing and sales company. As a financial database, it is designed to support various analytical queries related to the company's operations, sales, and financial performance. This (imaginary) company operates mainly in China, the United States, and Europe. Therefore the database is bilingual, with both English and Chinese values, and uses three currencies: CNY, USD, and EUR.

Specifically, there are 6 key dimensions: Period (time, monthly), Product, Region, Customer, Dealer, and Report Item (revenues and expenses). Also, there are extra dimensions including: Contract, Project, Currency, and Caliber.

The minimal granularity of the data is a payment of a project. Each project happens between APEX and a customer at a specific region, with an optional dealer, over a period of time. Payments of a project can be distributed over multiple months in that time period. Each project may contain multiple products. Each contract may contain multiple projects.

RubikBench contains 20 tables of 4 major categories, all of them are aggregated views over the fact table (which is not exposed directly):

• The INCOME tables, which contain data aggregated over project and dealer, and only includes revenues. The INCOME tables are the smallest tables in RubikBench, aiming for quick analytical queries.
• The BUDGET_AND_FORECAST tables, which contain data aggregated over project, customer, and dealer. These tables contain only amt/forecast/budget/target values. Notice that the semantics of these values could be counter-intuitive: While the target value is the target for monthly revenues and expenses in YYYYMM, as one would expect; the forecast value of YYYYMM is the forecast of YYYY12 (yearly) based on the information available at the end of YYYYMM; the budget of YYYYMM is a constant value indicating the yearly budget duplicated for each month in the year.
• The PROFIT_AND_LOSS table, which contains data aggregated only over project. It contains the most comprehensive dimensions and measures, including both revenues and expenses. It is the largest table in RubikBench, aiming to support detailed financial analysis.
• The SALES_LEDGER table, which contains the lowest granularity data, i.e., payment-level data. It is designed to support detailed audit and traceability of sales. However, it is limited to sales-related revenues and expenses only.

The products of APEX are divided into 3 major divisions: Automobiles, Accessories, and Services.

• Automobiles are produced by enterprise brand groups, which are 6 sub-brands under APEX. Each brand group has its own INCOME and BUDGET_AND_FORECAST tables.
• Accessories and Services each have their own INCOME and BUDGET_AND_FORECAST tables. Accessories only have equipment revenues and costs, while services only have service revenues and costs.

The default Caliber is code A, which clearly separates equipment and service report items, reflecting the real financial statistics. However, to facilitate the cooperation between different divisions, there are also Caliber B, which moves 5% of service revenue to equipment revenue.

Notice that due to historical reasons as well as query efficiency expectations, currencies and calibers are organized differently across different tables. For example, for the INCOME and PROFIT_AND_LOSS tables, different currencies and calibers are stored in different columns as column name suffixes (e.g., _cny_a, _usd_a, _eur_b, etc.); while for the SALES_LEDGER and BUDGET_AND_FORECAST tables, caliber and currency are separate columns that MUST be filtered on in the query predicates to avoid duplicated results.

Financial amounts presents both ptd (monthly) values and ytd (year-to-date) values. For example, ptd of YYYYMM means the amount for that month, while ytd of YYYYMM means the cumulative amount from YYYY01 to YYYYMM (inclusive). Furthermore, _py columns contain previous year data, which means that, for exammple, ytd py columns with period='YYYYMM' contain cumulative amounts from YYYY-1 01 to YYYY-1 MM, etc.

Query Statistics

RubikBench v0.9.2 contains 3198 fully human-verified queries covering diverse financial analysis scenarios, with comprehensive annotations including difficulty levels, query tags, and user context profiles. The queries are written in both English (1,689) and Chinese (1,509), reflecting the bilingual nature of the database.

3.1 Difficulty Distribution

Difficulty	Count	%	Avg. SQL Length	Min SQL Length	Max SQL Length
Simple	1,047	32.7%	218 chars	11 chars	1,240 chars
Moderate	1,531	47.9%	641 chars	213 chars	3,492 chars
Challenging	415	13.0%	1,755 chars	339 chars	3,635 chars
Nightmare	205	6.4%	1,881 chars	579 chars	4,183 chars

3.2 Table Coverage

Queries span all 4 major table categories (a query may reference multiple tables):

Table Category	Queries	Individual Tables
INCOME_*	1,615	INCOME_ALL (581), INCOME_WEYMOTCOR (153), INCOME_SHIMOTCOR (142), INCOME_TYRAUTGRO (132), INCOME_ACCESSORY (130), INCOME_NOVDYN (122), INCOME_SERVICE (121), INCOME_GALMOTGRO (120), INCOME_CYBVEHSYS (118)
BUDGET_AND_FORECAST_*	724	BUDGET_AND_FORECAST_ALL (124), BUDGET_AND_FORECAST_WEYMOTCOR (97), BUDGET_AND_FORECAST_TYRAUTGRO (85), BUDGET_AND_FORECAST_SERVICE (81), BUDGET_AND_FORECAST_SHIMOTCOR (79), BUDGET_AND_FORECAST_NOVDYN (77), BUDGET_AND_FORECAST_ACCESSORY (72), BUDGET_AND_FORECAST_CYBVEHSYS (54), BUDGET_AND_FORECAST_GALMOTGRO (53)
PROFIT_AND_LOSS	408	—
SALES_LEDGER	367	—

3.3 SQL Feature Distribution

SQL Feature	Queries	%
GROUP BY	2,296	71.8%
ORDER BY	2,139	66.9%
CASE WHEN	1,409	44.1%
COALESCE	876	27.4%
Subquery	768	24.0%
Window Function (OVER)	746	23.3%
LIMIT	635	19.9%
CTE (WITH)	438	13.7%
HAVING	334	10.4%
DISTINCT	260	8.1%
JOIN	94	2.9%
CAST	86	2.7%
UNION	28	0.9%

3.4 Query Occupations

Each query includes a simulated user profile with an occupation field, representing different roles that would typically interact with financial data:

Occupation	Count	%	Description
Sales	707	22.1%	Sales representatives and managers
Finance	675	21.1%	Financial analysts and accountants
Management	559	17.5%	Executive and strategic decision-makers
Guest	363	11.4%	External users with limited access
Developer	356	11.1%	Technical users and system developers
Unspecified	538	16.8%	No specific occupation context

Update Log

• [2025-02] RubikBench v0.9.2:

  • Added 3198 queries and query statistics (schema-grounded queries to be added).
  • Integrated BirdSQL MINIDEV benchmark and KaggleDBQA benchmark into RubikBench CLI.
  • Released RubikBench evaluation toolkit as open-source Python package: rubikbench.

• [2026-02] Updated RubikBench v0.9.1 database on huggingface: RubikBench.

  • Add support for .parquet format.
  • More realistic business logic, reduced the number of products per project
  • Diversify between SALES_LEDGER and PROFIT_AND_LOSS
  • Fixed some geographical region errors
  • Re-generate all projects and contracts
  • Expand to 9 dealers
  • ytd/ptd fix

• [2026-01] Initial release of RubikBench v0.9 database on huggingface: RubikBench.