arXiv CSopen access

TabQueryBench: A Query-Centric Benchmark for Synthetic Tabular Data

TabQueryBench: A Query-Centric Benchmark for Synthetic Tabular Data Jialin Zhang†,¶ Fenghao Dong‡ Yajie Zhou§ Vyas Sekar‡ Shinan Liu† † University of Hong Kong

¶ Tongji University

‡ Carnegie Mellon University

§ University of Maryland, College Park

[email protected], {fenghaod, vsekar}@andrew.cmu.edu, [email protected], [email protected]

arXiv:2607.03926v1 [cs.DB] 4 Jul 2026

ABSTRACT Synthetic tabular data support use cases like data sharing, model development under access restrictions, and rapid prototyping of analytical workflows. Modern generative models are evaluated by their statistical similarity, correlation structure, privacy, and downstream machine-learning utility. However, such evaluations leave a gap: they rarely test the structure that matters for analytical queries. We present TabQueryBench1 , a query-centric benchmark that uses SQL-shaped analytical queries as structural assessors for synthetic data fidelity. It provides an extensible foundation for query-centric synthetic-data evaluation. From 12 public sources of analytical queries, TabQueryBench taxonomizes recurring crossdomain logic into 44 reusable query templates and grounds them to each dataset via a policy-guided template-to-SQL pipeline. This makes queries schema-aware while preserving comparability across generative models. Across 49 datasets and 11 generative models, it activates 10–12 templates per dataset, producing more than 100 executable SQL queries per dataset. Our systematic experiments show five main patterns. First, current tabular generative models can have good distance-based fidelity, but they still fall short on query-centric fidelity: RealTabFormer achieves the highest query-centric fidelity, but it only reaches 0.75 ± 0.15 (REAL data score is 1.00). Second, tabular generative models struggle with very high-cardinality discrete support. Third, SOTA generative models preserve good global conditional query-centric fidelity, but fail more on local queries. Fourth, tail fidelity deteriorates as queries move toward the extreme tail; even the best generative model recovers only about 40.7% of real rare values. Finally, there is a fidelity-cost tradeoff in tabular data generation: BayesNet offers the strongest tradeoff, with slightly lower query-centric fidelity but much lower generation cost.

1

INTRODUCTION

Synthetic tabular data are now used for data sharing [5, 30, 46], model development under access restrictions [5, 8, 46], and rapid algorithm or system prototyping [21, 36, 45, 69]. As these use cases mature, benchmarking matters more, because what makes synthetic data useful in these settings is not distributional resemblance alone, but the preservation of the analytical properties that the database community has long treated as the core measure of data quality [9, 14, 31, 42, 48, 50, 65, 66]. Synthetic data consumers often care less about synthetic records as standalone samples. They care more about the functions that those records make possible. A useful synthetic dataset does not merely approximate marginal distributions. It should ideally expose a 1 TabQueryBench is open-sourced at https://github.com/TabQueryBench/ TabQueryBench and https://huggingface.co/datasets/TabQueryBench2026/ TabQueryBench/tree/main

controlled approximation of hidden dataset attributes, such as table structure, column relationships, valid joins, and executable queries [36, 45, 69]. This structural view is important for (1) collaboration, where a partner wants to preview the structure and quality of a sensitive dataset before accessing the real data [5, 46]; (2) software and data-system testing, where engineers need realistic schemas, constraints, edge cases, and relational consistency [5, 45, 46, 69]; and (3) query-centric analytics, where the target is not an individual row but the answer to aggregates, joins, SQL functions, and textto-SQL workloads [7, 36]. While recent libraries and benchmarks have made comparison more systematic [13, 16, 18, 24, 39, 51–53, 56, 61, 62], most evaluations still focus on column-level distributional scores, one-shot predictive utility, and privacy reports. At a high level, there exists a disconnect between the evaluation methods used by current tabular generative models and the types of structural information that practitioners actually care about. We find that picking generative models based on conventional distance-based metrics may even be misleading for practitioners [26, 64, 68]. Our key observation is that existing synthetic tabular benchmarks have limited coverage of the analytical structures that occur in real analytical queries. Current benchmarks score column-level distance, predictive ML utility, or task-specific utility proxies [17, 25, 34, 39, 57], none of which directly assesses the analytical properties that diverse downstream workflows care about. A useful contrast comes from the database community, which has long treated occurring query sets as first-class evaluation objects: benchmarks such as TPC-H and TPC-DS record ocurring decision-support SQL queries from real OLAP practice [14, 42, 48, 50]. Although designed to compare database engines, such recurring query patterns also provide natural structural assessors for synthetic data. In contrast, existing synthetic-data benchmarks do not yet use them in this role, which shows a natural reality/generation target mismatch. Our answer: TabQueryBench. We argue for a shift from the conventional benchmarking used for tabular generative models to query-centric benchmarking, where domain-relevant queries serve as structural assessors of synthetic-data fidelity. The key design challenge is to make these assessors both realistic and generic: real analytical questions are tied to dataset semantics, but a benchmark must evaluate many datasets, tasks, and generative models using queries built from a common set of templates. TabQueryBench addresses this challenge with a two-stage construction pipeline. First, it taxonomizes recurring analytical logic from public sources of analytical queries into reusable five-family templates (i.e., subgroup, conditional, tail/rarity, cardinality/range, and missingness). Second, it grounds these templates to each benchmark dataset through a constrained realization policy that binds template roles to schemaappropriate columns and validates the resulting queries. This design

TabQueryBench: A Query-Centric Benchmark for Synthetic Tabular Data RTF

2

BayesNet

0.748 0.697

Template Library

5

TabPFGen

0.639

Reusable templates grouped by analytical intent

6

CTGAN

0.623

1

Subgroup Conditional Tail Cardinality Missingness 44 templates

SQL Query

2

From profile to executable grounded SQL

Profile

Select

Generate SQL

7

TVAE

0.591

8

ForestDiff

0.589

9

TabbyFlow

10

TabSyn

0.555

11

TabDDPM

0.543

l

0.644

ona

0.655

TabDiff

diti

ARF

Con

3 4

y

Docs

alit

SQL Repos

1

Car din

Benchmark Queries

Subgroup

Average Ranking

Workload Source

0.589

Mi ss

ing

Query-Centric Fidelity Subgroup

100 80 60 40 20

67 68 63

Conditional

61

ss

Missingness

Tail / Rarity

Cardinality / Range

97 95 96 95 97 94 96 95 96 94 94 79

76 59 55 63 63 57 61

64 50 52 47

41

44 43

46 48

il Ta

ne

90

89

80

43 43

35

85 84

76 58

61

65

75

45

43 24 8

19

27 13

9

18 21

0

Figure 1: Overview of TabQueryBench, including the benchmark design and representative evaluation results. preserves schema-level realism while enabling comparable evaluthat appear in the real table. In the discrete-tail diagnostic, even ation across datasets, domains, and generative models (Figure 1). the best generative model recovers only about 40.7% of the real TabQueryBench also provides fine-grained feedback on specific rare values. This result shows that rare-event queries remain queries and query families rather than only an aggregate score. fragile under synthetic data. This gives both generative-model developers and practitioners ac• There is a clear cost-fidelity tradeoff in tabular data gentionable signals for tuning, debugging, and selecting generative eration and BayesNet achieves the best balance. RealTabmodels. For this query construction and grounding pipeline, we Former achieves the best query-centric fidelity, but it is also demonstrate that the query-centric conclusions are broadly stable much slower than the other generative models. BayesNet gives across multiple SQL-regeneration runs. slightly lower fidelity, but it runs orders of magnitude faster. This In this work, we identify the following key findings on our benchmakes BayesNet the more practical choice when users care about mark of 49 datasets and 11 tabular generative models: both query quality and generation cost. • Current tabular generative models can have good distancebased fidelity, but they still fall short on query-centric fidelity. RealTabFormer achieves the highest query-centric fidelity, while BayesNet offers the strongest fidelity-cost tradeoff. Yet even these two models remain substantially below the REAL reference. This gap is especially clear when compared with their stronger performance under distance-based fidelity metrics.

TabQueryBench is released as an open benchmark. The code, template catalog, and query-generation artifacts are available on GitHub (https://github.com/TabQueryBench/TabQueryBench), and the benchmark data package and docker images for generative models are mirrored on Hugging Face (https://huggingface.co/datasets/ TabQueryBench2026/TabQueryBench/tree/main). The release is intended to make the benchmark inspectable and reusable.

• Tabular generative models fail to synthesize columns with very high-cardinality discrete support. When a column contains hundreds or thousands of distinct values, many generative models preserve only a fraction of the real support. For example, on dataset c18, the real title column has 96,777 distinct values, while BayesNet generates only 242 and CTGAN generates 19,630.

2

BACKGROUND AND MOTIVATION

In this section, we present backgrounds of tabular generative models, previous related benchmark designs, and highlight the misalignment of focuses between synthetic data evaluation and data analytics.

• SOTA generative models preserve good global conditional query-centric fidelity, but fail more on local queries. When a conditional query adds only a filter, fidelity often drops even though the grouping keys and aggregation logic stay unchanged. For example, on dataset c6, TabDiff scores 0.632 on the global 2D count surface but only 0.276 after the local filter is added.

2.1

Tabular Generative Models

Tabular synthesis learns a distribution over records with numerical and categorical fields. It supports privacy-preserving data sharing, data augmentation, imputation, and benchmarking when real data is sensitive or scarce. Tables differ from images and text because each column has its own structural type and meaning [15, 73, 75]. A useful generative model should preserve valid values and rare

• Tail fidelity deteriorates as queries move toward the extreme tail, even for SOTA generative models. Under stricter rarity thresholds, generative models increasingly lose rare values 2

Benchmark

Target setting

Evaluation Focus

#Datasets

#Models

Subgroup

Cond.

Tail

Miss.

Highcard.

Dist.

Cost

Synthcity [51] Data-centric [24] SDGym [13] SDNist [53, 62] SyntheRela [27] SynthRO [52] SynthEval [39] Multi-dim. [56] DP benchmark [61] TabArena [18] TabStruct [32]

Generic tabular Generic tabular Generic tabular Release task Multi-table Health domain Generic tabular Mixed data Private tabular Tabular ML Generic tabular

Metric suite Global fidelity Global fidelity Challenge score Relational fidelity Domain validation Metric suite Global fidelity Privacy–utility Predictive utility Structural metrics

18+u 11 22+u 4 39

5+u 5

✗ ✗ ✗ ✗ ✗ ✗ ✗

✗ ✗ ✓ ✗

7 51 29

12 16 13

△

✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

✓ ✓ ✓ ✗

u u u

✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

△

u u u

✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

TabQueryBench (ours)

Generic tabular

Query-centric fidelity

49

11

✓

✓

✓

u u 6

✗ ✗ ✗ ✗ ✗ ✗ ✓ ✗

△

△ ✗ ✗ ✗ ✓ ✓ ✓

✗

✗

✓ ✓ ✓ ✓ ✗ ✓

✓

✓

✓

△

△ ✗

△

△

Table 1: Comparison of related benchmarks for synthetic tabular data and adjacent tabular evaluation tasks. We report each benchmark’s target setting, primary evaluation object, scale, and coverage of five query-centric property families, classical distance, and cost. ✓ = direct, △ = partial, ✗ = absent, u = user-specified. categories. It should also respect column constraints and crosscolumn dependencies. Recent models differ mainly in their inductive biases. Bayesian models such as BayesNet use an explicit graphical factorization, which gives a compact view of low-order dependencies [47]. Treebased methods such as Adversarial Random Forests (ARF) use adaptive partitions of the feature space and match the local structure of tables well [71]. Neural models then shifted the field toward learned representations. CTGAN and TVAE handle mixed feature types through conditional generation and latent-variable modeling [72]. Diffusion methods such as TabDDPM, TabSyn, TabDiff, and ForestDiffusion replace adversarial or reconstruction losses with iterative denoising [33, 38, 55, 74]. Transformer-based models such as REaLTabFormer and TabPFGen use attention or pretrained tabular priors to capture broader dependencies across columns and records [40, 60]. Flow-matching models such as TabbyFlow learn a transport path from noise to data and provide a newer alternative to diffusion-based sampling [22]. Overall, the trend moves from explicit structures to flexible neural samplers, while the central goal remains the same: preserve useful tabular structure without copying the training data.

2.2

rather than tabular generative models [18, 27, 45, 52, 56, 61]. Table 1 summarizes each benchmark’s primary evaluation object, reported scale, and coverage of the five query-centric property families together with classical distance and cost baselines. These resources are highly useful, but they do not make the answers to analytical queries the main object of evaluation. Distributionand ML-based benchmarks can report strong global fidelity even when synthetic data gives wrong answers for filtered subgroups, conditional slices, rare rows, or missingness-dependent queries [3, 29, 39, 57, 61]. TabQueryBench targets this gap by evaluating whether synthetic data preserves the answers induced by reusable SQL-shaped analytical queries (see Section 2.4 for concrete mismatch examples). Rather than treating query behavior as one additional metric, TabQueryBench uses dataset-grounded analytical queries as structural assessors for synthetic tabular data, and Table 1 highlights this shift by separating query-family coverage from distance and cost coverage.

2.3

Database Query Benchmarks

The database community provides a useful contrast and sources of analytical query patterns: it has long treated recurring query sets, not samples alone, as the object of evaluation. TPC-H and TPC-DS define decision-support benchmarks with generated schemas, parameterized business queries, and controlled execution rules [42, 48, 65, 66]. Recent public resources extend the same idea to modern analytical settings: ClickBench uses an anonymized web-analytics dataset and a reproducible collection of SQL queries; H2O’s dbbenchmark compares data-processing systems through databaselike operations such as grouping and joins; and RTABench targets real-time application analytics with normalized schemas, selective filters, joins, and pre-aggregated views [11, 23, 63]. These benchmarks evaluate database engines rather than tabular generative models. TabQueryBench repurposes their central idea: analytically meaningful SQL patterns can serve as reusable assessors, and synthetic data should preserve the answers those assessors induce.

Tabular Generative Model Benchmarks

Existing benchmarks for synthetic tabular data differ along two axes: the setting they target and what they directly evaluate. These two dimensions largely determine what conclusions a benchmark can support and what aspects of synthetic data quality it emphasizes. Most general-purpose benchmarks target single-table generation and score global fidelity, privacy, cost, or downstream machinelearning utility. Synthcity and SynthEval provide reusable evaluation infrastructures with user-specified datasets, generative models, and metrics [39, 51]. SDGym and the data-centric benchmark of Hansen et al. instead define fixed dataset suites and standardized protocols for controlled head-to-head comparison [13, 24]. Challenge-style resources such as SDNist are more task-specific: they define public release scenarios and scoring rules for a shared leaderboard [53, 62]. Other benchmarks specialize in narrower settings. SyntheRela studies multi-table synthesis, SynthRO focuses on health-oriented validation, Sidorenko et al. evaluate mixed and contextual data, Tao et al. benchmark differentially private mechanisms, and TabArena evaluates tabular machine-learning systems

2.4

Mismatch Between the Distance-based and Query-centric Fidelity

Figure 2 motivates that distance-based fidelity alone is insufficient for evaluating synthetic data. It can diverge with query 3

N6 Query-centric Fidelity

C9

M4

1

3.1

6

2

4

0.8

5

1 2

0.6

4

6

5

5 3

0.4 6

1

3 3

0.2

4 2

0

0

0.01 0.02 0.03 0.04

0

Wasserstein 1 BayesNet 4 TPF

0.2

0.4

0.6

0.8

0

0.2

JSD

0.4

0.6

Principle 2: Decomposable and diagnostic. TabQueryBench supports queries that capture both global and local structures within datasets, with different queries focusing on different structural properties. These structural assessors provide more informative signals than distance-based fidelity metrics, which often compress quality into coarse scores over entire rows or columns. By making evaluation decomposable, TabQueryBench enables quality attribution at finer granularity and helps diagnose where structural fidelity is preserved or degraded.

0.8

JSD

2 CTGAN

3 T-DDPM

5 T-Syn

6 TVAE

Figure 2: Comparison between query-centric fidelity vs. distance based metrics. The x-axis is distance-based fidelity (Wasserstein for N6; JSD for M4 and C9), and the y-axis is a query-centric fidelity score (the higher the better with 1 being the best, queries are related to conditional structures). Numbers 1–6 identify the same models across figures.

Principle 3: Reusable by construction. TabQueryBench is designed as a reusable pipeline that works across datasets without requiring dataset-specific redesign. Its query generation, execution, and evaluation components are modular and can accommodate new tables, schemas, and domains. This construction makes TabQueryBench broadly applicable, extensible, and easy to plug into different data settings.

because a query often depends on a small conditional structure (details in Section 4.2) rather than the whole distribution. We use three dataset-query pairs to illustrate three mismatch patterns across numerical (N), mixed (M), and categorical (C) datasets. In N6, an EEG-based epilepsy classification dataset, the query asks: within each class y, which X11 groups contribute the largest share of X10? This is a class-conditional ranking query. TabPFGen has one of the smallest Wasserstein distances, but it shifts the dominant groups away from the real high-impact buckets, such as (𝑦 = 3, 𝑋 11 ∈ {4, 5, 6, 2, 3}), toward mostly negative X11 buckets. Thus, the distance-based fidelity is high, but the query-centric fidelity is low. TabSyn has a similar Wasserstein distance, but it preserves more of the real top groups and therefore achieves higher query-centric fidelity. In M4, a medical insurance dataset, the query asks: which (smoker, age) groups contribute the largest share of total children? This example shows that a model can look good under a distance-based metric while changing the conditional support used by the query. TVAE achieves a low JSD, but it over-concentrates children around 1 and compresses the (smoker, age) support. As a result, the grouped aggregate becomes dominated by a narrow age range. TabPFGen preserves both the children distribution and the smoker-age support better, so it performs well under both metrics. In C9, the Amazon Employee Access Challenge dataset, the query asks: for a target ROLE_FAMILY, are the relevant ROLE_TITLE keys preserved for rate analysis? This case shows the reverse pattern. TabSyn and CTGAN have worse JSD because they distort the broader ROLE_TITLE–ROLE_FAMILY structure. However, they still retain many query-relevant ROLE_TITLE keys, so their query-centric fidelity remains high. These examples show that distance-based fidelity and querycentric fidelity measure different aspects of synthetic data quality. Reliable evaluation therefore needs both global metrics and diverse downstream queries.

3

Design Principles

Principle 1: Grounded in real analytical queries. TabQueryBench is grounded in analytical queries drawn from real-world practice. Rather than relying on synthetic or arbitrary assessors, it characterizes datasets through the ways users actually interact with tabular data. This makes the evaluation representative of practical analytical use cases and aligned with downstream analytical needs.

1

3.2

Scope

TabQueryBench is scoped along two dimensions: the queries it considers and the generative models it evaluates. Query scope. The benchmark focuses on reusable analytical query patterns that recur across public OLAP and decision-support query collections. In scope are cross-dataset patterns such as subgroup comparison, filtered conditional dependence, tail and rarity analysis, cardinality and range structure, and missingness behavior [42, 48, 49]. These patterns are intended to capture structural properties that transfer across datasets and domains, while still being grounded to each dataset through schema-aware realization. Because of the characteristics of synthetic tabular generative models, out of scope are queries whose semantics are inherently non-transferable or not central to analytical fidelity: point lookups by primary key, forensic debugging queries that depend on specific row identifiers, private business-rule queries tied to a single organization, temporal forecasting beyond simple time-bucket aggregates, and long relational join chains. The current design is most mature for single-table settings, with relational coverage limited to two-table joins as described in Section 4.3. Generative model scope. TabQueryBench evaluates mimicrystyle synthetic-data generation. Each generative model observes a real training split and produces a synthetic table or database under the same schema, with the objective of preserving the analytical structure of the original data. This scope matches common synthetic-data release and benchmarking settings, where the central question is whether the generated data supports the same kinds of analysis as the real data. We therefore do not evaluate differentially private synthesis, where utility is explicitly traded against

BENCHMARK SCOPE

In this section, we present the scope of TabQueryBench. 4

Bayesian

Tree

GAN

VAE

Diffusion

Transformer

Flow-matching

BayesNet [47]

ARF [71]

CTGAN [72]

TVAE [72]

TabDDPM [38]; TabSyn [74] TabDiff [55]; ForestDiffusion [33]

REaLTabFormer [60] TabPFGen [40]

TabbyFlow [22]

Note. Throughout the paper, abbreviated model names are used in tables and figures for compactness. BayesNet=Bayes, TabDDPM=T-DDPM, TabSyn=T-Syn, TabDiff=T-Diff, ForestDiffusion=F-Diff, REaLTabFormer=RTF, TabPFGen=TPF, and TabbyFlow=T-Flow

Table 2: Generative model suite grouped by core modeling technology.

TabQueryBench: Query Taxonomy 1

Subgroup

We collect our 49 datasets from four public tabular-data sources: UCI [37], Kaggle [35], OpenML [44], and HuggingFace [28]. All 49 datasets are single-table tabular datasets. The suite covers a broad range of domains, including business and operations, healthcare and insurance, education and public policy, industrial and engineering settings, media and consumer content, and scientific or life-science data. This broad coverage helps the benchmark reflect the variety of tables that real users analyze. Another criterion is that we choose datasets that can stress-test synthetic data in different ways. Row counts range from 1.5K to 2.46M. Column counts range from 3 to 1,559. In addition, 10 datasets contain high-cardinality features (more than 50 distinct values in a column), and 12 contain substantial missingness. Some datasets are especially useful as stress cases, such as Internet-Advertisements as a very wide table, SECOM as a high-dimensional dataset with heavy missingness, and US Census Data 1990 as a very large table.

(16 templates)

Structure

(13 templates)

aggregated subgroup patterns

Size

(3 templates)

absolute group counts

Representative query idea Compare average charges across age groups, or rank regions by customer coverage.

2

Conditional

(13 templates)

Global / grouped

(8 templates)

conditional signal by group

Local / filtered

3

Tail / Rarity

(8 templates)

Tail Coverage

(4 templates)

recover rare members or outliers

(5 templates)

Tail Size

(4 templates)

same summary after slicing

tail mass, rarity, or level

Representative query idea

Representative query idea

Within each category, ask what fraction satisfies a condition before a local slice.

Identify the rarest users or unusual rows, or estimate how much mass lies in the tail.

4

Missingness

Marginal Missingness

(3 templates)

(1 templates)

overall missing-rate signal

Broad Co-Missingness

5

3.4

We evaluate 11 tabular generative models spanning the major modeling families currently used for synthetic tabular data. Table 2 groups them by their core modeling technologies: a Bayesian network baseline [47], a tree-based generative model [71], adversarial and variational neural generative models [72], diffusion-based methods [33, 38, 55, 74], transformer-based generative models [40, 60], and a recent flow-matching approach [22]. This breadth is intended to make the benchmark representative of the current methodological landscape rather than of a single modeling paradigm. For each dataset, generative models are run in the same synthetic setting: the generative model is trained on the real training split and then asked to produce a synthetic table under the original schema. Keeping the training data, schema, and output contract fixed allows differences in query-centric fidelity to be attributed to generative-model behavior rather than to differences in task formulation. Throughout the paper, we use the compact model names listed in Table 2 when reporting results. Generative models that fail to finish, exceed resource caps, or emit invalid outputs are recorded explicitly rather than silently dropped: instability is itself part of practical model quality [6, 41].

Cardinality / Range (4 templates)

Discrete

(3 templates)

support, share, and high-card keys

(2 templates)

Continuous

(1 templates)

missingness by state or bucket

range-envelope preservation

Representative query idea

Representative query idea

Track how missing-rate changes across states, buckets, or grouped contexts.

Check whether discrete support counts or a numeric range envelope survives generation.

Figure 3: The Query Template Taxonomy. a privacy budget; conditional generation under task-specific constraints, where query-centric fidelity is not the primary target; or text-to-table generation from natural-language prompts. The generative models included in the current roster all fall within this mimicry-style setting.

4 3.3

Generative Model Suite

Dataset Suite

TabQueryBench DESIGN

The core of TabQueryBench is the construction of dataset-grounded analytical queries. This process proceeds in two stages. Template construction (Stage 1, Section 4.2) distills public sources of analytical queries into a reusable template library, fixed across datasets. Query generation (Stage 2, Section 4.3) grounds each template to an individual dataset by deciding which schema columns realize each template role, then expanding the template into executable SQL. Lower-level implementation details, including runtime traces and the per-template policy bundles, are deferred to Appendix A.

TabQueryBench curates 49 datasets organized by feature regime: 19 categorical-dominant datasets, 19 numerical-dominant datasets, and 11 mixed-type datasets. Categorical datasets often test discrete support coverage, subgroup identity, and rare-state membership, while numerical datasets more often test range structure, continuous conditional behavior, and tail mass geometry. Mixed datasets test the coupling of these two regimes [20]. Per-dataset metadata (rows, columns, and source) is provided in Appendix C, Tables 11. 5

4.1

Pipeline Overview

Question Grounded SQL Find subgroup counts and rank groups by size.

Which buying groups are most frequent?

Subgroup

template S-2.1 dataset c2

SELECT buying [<group_field>], COUNT(*) AS row_count FROM c2 [<table>] GROUP BY buying [<group_field>] ORDER BY row_count DESC;

Count rows per (𝑥, 𝑦) cell inside a local slice.

4.2

Grounded Question

Family

Stages 1 and 2 separate two responsibilities that are easily conflated. Stage 1 is dataset-agnostic: it produces query templates whose analytical shape is fixed by public evidence from analytical queries. Stage 2 is dataset-specific: it produces concrete SQL queries by adapting each template to the schema and data profile of one dataset, under a constrained generation procedure that is not free-form prompt authoring. Table 3 previews the grounded questions and SQL that Stage 2 produces.

Stage 1: Template Construction from Public Query Sources

Within rows where Subtopic equals Linear Transformations, how many rows fall into each Student Country and Question Level combination?

Conditional

Source registry. The current template library is constructed from twelve public sources organized into three categories. The first category is public benchmark material: TPC-H qgen [66]; TPCDS qualification and Altinity repositories [1, 12]; ClickBench [11]; H2O db-benchmark [23]; and RTABench [63]. The second category is curated public SQL repositories: Exploratory-Analysis-of-CarEvaluation-Dataset-with-SQL [43], Insurance-Cost-Project-UsingSQL [4], and insurance-sql-analysis [54]. The third category is from official database engine documentation: BigQuery approximate aggregate functions [19], ClickHouse aggregate functions [10], and Apache Druid SQL functions [2]. The benchmark-facing sources inherit the database community’s long-standing practice of encoding analytical behavior through reusable query templates rather than ad hoc one-off queries [42, 48, 49]. The full source-to-template attribution is given in Appendix A, Table 7. Some templates may originate from multiple benchmark sources; we deduplicate overlapping templates so that recurrence does not inflate coverage. Why these sources? We use these sources because they can provide public provenances, and can be grounded in analytical query sets used in real database and analytics systems. They give us query patterns that practitioners and benchmark designers already measure, rather than invented prompts. SQL repositories and engine documentation add realistic single-table analyst queries and commonly supported aggregate idioms.

template C-L1 dataset c6

SELECT "Student Country" [<group_x>], "Question Level" [<group_y>], COUNT(*) AS row_count FROM "c6" [<table>] WHERE "Subtopic" = ’Linear Transformations’ [<slice_predicate>] GROUP BY "Student Country" [<group_x>], "Question Level" [<group_y>] ORDER BY row_count DESC;

Count rows by group Among the top 3% of balance, inside an upper-tail slice. which previous values occur most often?

Tail / rarity

template T-2.2 dataset m8

Missingness

template M-2.1 dataset c5

Cardinality / range

template K-1.1 dataset c18

Query family taxonomy. A family is a class of SQL query shapes characterized by a shared aggregation primitive and a shared analytical question. We identify five families that recur across the source registry for single tables. They each test analytically distinct properties of synthetic data. The full taxonomy is shown in Figure 3, with 44 templates in total, organized as follows: • Subgroup contains 16 templates split into Structure (13) and Size (3). The Structure branch keeps the subgroup object fixed but varies the grouped summary signal, including totals, means, distinct coverage, robust summaries, and winner-style aggregate views. The Size branch instead asks only how large each subgroup is, using absolute count or support views over onedimensional and two-dimensional groupings. • Conditional contains 13 templates split into Global structure (8) and Local slices (5). The global branch keeps the conditional summary at full grouped scope, whereas the local branch reuses the same grouped conditional scaffold after adding a predicatedefined slice.

SELECT previous [<group_field>], COUNT(*) AS row_count FROM m8 [<table>] WHERE balance > PERCENTILE(balance, 0.97) [<tail_predicate>] GROUP BY previous [<group_field>];

Compute a target missing rate for each state.

For each odor category, what fraction of rows has missing stalk-root?

SELECT "odor" [<state_col>], AVG(CASE WHEN "stalk-root" [<target_col>] IS NULL THEN 1.0 ELSE 0.0 END) AS missing_rate FROM "c5" [<table>] GROUP BY "odor" [<state_col>] ORDER BY missing_rate DESC;

Enumerate a high-cardinality attribute by support.

How many rows are associated with each wine title?

SELECT title [<value_col>] AS category_value, COUNT(*) AS support FROM c18 [<table>] WHERE title [<value_col>] IS NOT NULL GROUP BY title [<value_col>] ORDER BY support DESC, category_value;

Table 3: One question and SQL example for each of the five query families. The blue text marks dataset-specific grounding, and the orange notes show which template slot each grounded SQL fragment instantiates. Details in Table 10. • Tail / rarity contains 8 templates split into Tail Coverage (4) and Tail Size (4). Coverage-oriented templates focus on whether the correct rare members or outliers are recovered, while sizeoriented templates focus on whether above-threshold magnitude or data size remains plausible. • Missingness contains 3 templates split into Marginal Missingness (1) and Broad Co-Missingness (2). We use missingness to describe the pattern of absent or null values, not only the overall count of missing cells. The marginal branch leaves the missingrate signal unconditional, whereas the broad co-missingness branch conditions on a bucket. 6

• Cardinality / range contains 4 templates split into Discrete (3) and Continuous (1). The discrete side tracks support/rank behavior over observed values, while the continuous side tracks min/max and range structure over numerical coverage. Table 3 gives one grounded SQL generation for each query family. Blue texts mark instantiated slots or grounded constants, while the uncolored SQL remains the fixed template shared across datasets.

profile or from the fixed template policy. Then the pipeline checks type compatibility, must-fix preservation, and executability, and rejects invalid generations. The SQL generator also generates the natural-language question, only as a readable interpretation of the accepted SQL generation. As illustrated in Figure 10, the pipeline maps each template to eligible dataset columns and produces a dataset-specific grounded problem together with its SQL generation. Only the SQL generator uses the LLM, and it does so under fixed template policies.

Why these query families? A query family enters the taxonomy only if it passes three tests. First, the family must isolate one structural property that no other family covers. Each of these properties might have very different failure modes to uncover. Second, the family must carry provenance: analysts already probe the property with recurring SQL. Every template traces to a public benchmark suite, SQL repository, or engine document. Third, the family must sit in a blind spot of the distance-based metrics. A synthetic table can match aggregate statistical distances while still giving wrong answers for filtered subgroups, conditional slices, rare rows, missingness-dependent queries, or high-cardinality support. These three requirements connect the taxonomy directly to the motivation for query-centric evaluation: each family acts as a structural assessor for a recurring analytical query pattern whose answer may not be preserved by distributional resemblance alone. We do not claim that the five families exhaust SQL. They exhaust the query-centric structural properties that are both recurring in public single-table analytical workloads and broadly groundable across heterogeneous tabular datasets.

4.3

4.4

Pipeline Statistics

TabQueryBench contains 44 templates (Appendix A, Table 10). Each dataset activates 10–12 templates on average, producing more than 100 executable SQL queries with associated natural-language questions per dataset (Appendix A). Across the 49-dataset suite and 11 generative models, this sums up to several thousands of benchmark queries and tens of thousands of generative-model– dataset–query evaluations.

5

EVALUATION

We holistically evaluate TabQueryBench and derive key findings from the results.

5.1

Settings

We evaluate TabQueryBench on 49 datasets and 11 synthetic tabular generative models. For each dataset, we split the real table into a training split and a held-out evaluation split (4:1). Each generative model is trained only on the training split and is then asked to synthesize a table under the same schema. The synthetic table is generated with the same number of rows as the evaluation split.

Stage 2: Dataset Grounding for Query Generation

Stage 2 grounds reusable templates to the 49 benchmark datasets by converting each template into executable SQL for each dataset where the template is applicable. The pipeline first reads the dataset schema and computes a lightweight dataset profile containing column types, missingness rates, support sizes for discrete columns, and numeric ranges for continuous columns. It then uses that dataset profile only to determine whether a template is admissible on the dataset. The profiling does not search for columns that make the resulting query visually interesting or artificially easy. The pipeline then maps each template to eligible dataset columns under the fixed policy of that template. A template may require a grouping column, a numeric measure, a predicate column, an ordering column, or a column used to define a bucket or threshold. The binding must satisfy the type and support requirements of the template. For example, a subgroup template needs a valid grouping key, a range template needs an ordered numeric column, and a missingness template needs a column whose missing values can be queried. If no eligible binding exists, the pipeline skips that template on the dataset. After the pipeline accepts a binding, the SQL generator expands the template skeleton under that binding. The SQL generator fills the selected columns into the aggregate expressions, predicates, WHERE clauses, GROUP BY clauses, ordering clauses, and limits specified by the template, producing the SQL generation used by the benchmark. When the template requires a constant, threshold, or bucket boundary, the SQL generator derives it from the dataset

SQL grounding setting. The profiler, binder, and validator are deterministic. Only the SQL realizer calls an LLM, and it does so only inside the constrained template-to-SQL realization step. We use the ChatGPT 5.4 API for that step, and the mean API cost is 5.40 per dataset. The validator rejects invalid SQL, incompatible realizations, and queries that fail execution, and the SQL realizer retries under the same fixed template policy. The full summary by generative model, including the common-9 runtime audit columns, is deferred to Appendix Table 8. Generative Models and Hyperparameter Settings. We package each tabular generative model in its own Docker image. For each generative model, we use the recommended configuration from the original implementation when available, and otherwise tune within a bounded search range chosen to balance quality and runtime. For each generative model, we tuned the main capacity, optimization, and training-budget parameters within a bounded search space. For ARF, we varied the number of trees from 10 to 150, the maximum number of iterations from 3 to 20, the minimum node size from 1 to 7, and the split tolerance delta from 0 to 0.02. For BayesNet, we varied the number of rows used for fitting and structure learning from 2K–120K and 1K–25K, respectively, together with the maximum number of discretization bins from 4 to 8 and the maximum categorical levels from 32 to 128. For CTGAN and TVAE, we tuned the number of epochs, batch size, embedding dimension, and hidden-layer dimensions: CTGAN used 50–200 epochs, batch 7

105

0.9

Synthetic distinct count

Query overall score

1

0.8 RTF

0.7 0.6 0.5 0.5

Bayes ARF T-Diff CTGAN TPF T-Flow F-Diff TVAE T-Syn T-DDPM

0.6 0.7 0.8 0.9 Distance-based overall score

Bayes F-Diff T-Flow TPF TVAE

102

103 104 Real distinct count

103

102

1

Figure 4: Distance-based and query overall scores do not align across generative models. Points below the diagonal have lower query-centric fidelity than distance-based fidelity; detailed values appear in Appendix Table 8.

105

Figure 5: Real vs. synthetic distinct counts on highcardinality discrete columns. Each point is one generativemodel–column pair from the discrete rows in Table 4. that matters for analytical queries. The full per-generative-model summary is deferred to Appendix Table 8. The appendix table also shows that the failure is not uniform across query families. Missingness is comparatively easy, with all generative models above 0.94, but Tail/Rarity and Conditional remain much harder. Even the best tail score is only 0.45±0.27 , and the best conditional score is only 0.64±0.27 , both achieved by RealTabFormer. Thus, distance-based fidelity can give an overly optimistic picture: it hides the subgroup, conditional, and rare-event failures that matter for real analytical use.

sizes of 10–128, embedding dimensions of 8–32, generator/discriminator hidden dimensions from (16, 16) to (64, 64), and pac values from 1 to 10; TVAE used 100–500 epochs, batch sizes of 256–500, embedding dimensions of 32–256, and encoder/decoder hidden dimensions from (64, 64) to (256, 256). For diffusion-based methods, ForestDiffusion varied n_t from 4 to 20, n_estimators from 5 to 100, duplicate_K from 2 to 20, max_depth from 3 to 6, and max_train_rows from 4,096 to 50,000; TabDDPM varied the number of diffusion timesteps from 200 to 1,000, training steps from 40 to 5,000, batch size from 64 to 1,000, and learning rate from 10−4 to 10−3 ; TabDiff varied the number of epochs from 100 to 1,000. For transformer- and flow-based generative models, RealTabFormer used 5–100 epochs, TabbyFlow used 100–700 epochs, and TabPFGen varied fit_max_rows from 512 to 4,096 and gen_chunk_rows from 64 to 256 while keeping its SGLD settings fixed. Finally, for TabSyn, we varied the VAE and diffusion training epochs from 3 to 20 and the VAE batch size from 16 to 256.

5.2

104

ARF CTGAN RTF T-Diff T-Syn

5.3

Where Do Failures Under Query-Centric Fidelity Concentrate?

A major goal for TabQueryBench is to localize failure modes rather than reporting only an aggregate query score. We study two complementary axes: query families, which reveal difficult structures to preserve, and data regimes, which reveal how categorical, numerical, and mixed schemas change the failure profile.

How Well Do State-of-the-Art Tabular Generative Models Perform under Query-Centric Evaluation?

5.3.1 Breaking Down By Query Family. We break down querycentric fidelity across the query families and demonstrate analysis and insights. We first localize failures by query family. This view asks which kinds of analytical query objects are most fragile under synthetic generation, rather than averaging all query behavior together. In the main text, we focus on three representative hard regimes. Conditional queries test whether grouped structure remains reliable after the analysis is restricted to a filtered local slice. Tail and rarity queries test whether rare regions remain both identifiable and queryable as the support becomes more extreme. Highcardinality queries test whether synthetic data preserves the large discrete support that matters for analytical queries. For each family, we follow the same progression: we first define the query object, then show one grounded example, then summarize the benchmark evidence, and finally explain what the result implies about current synthetic tabular generative models.

Finding 1: Current state-of-the-art tabular generative models often appear faithful under distance-based metrics, but still fall well short on query-centric fidelity (Figure 4). To provide an overview of TabQueryBench results, we evaluate whether current tabular generative models preserve the analytical behavior exposed by dataset-grounded analytical queries. We are also curious about whether this query-centric approach provides a different view compared to conventional distance-based fidelity. Figure 4 compares each generative model’s distance-based overall score with its query overall score. Every generative model lies below the diagonal, so the query-centric view is uniformly stricter than the distance-based one. The ordering also shifts. RealTabFormer remains strongest overall, but ARF and TabPFGen look especially strong under distance-based fidelity relative to their queryside scores, while BayesNet remains comparatively stronger once we ask whether the same synthetic tables preserve the structure

Finding 2: Current SOTA tabular generative models often fail to preserve high-cardinality discrete support (Figure 5 and Table 4). 8

Dataset Rows

Column Name

Kind

Real ARF

Bayes CTGAN F-Diff

c14 c15 c17 c17 c18 c18 c18 c3 c3 c19

7 nom_9 listed_in date_added region_1 title winery ATRINS-DONOR-521 CCAGC channel_title

discrete discrete discrete discrete discrete discrete discrete discrete discrete discrete

1219 2218 484 1634 1179 96777 15786 2541 2426 2181

64 2216 421 1375 255 242 256 1572 1562 256

300,000 600,000 8,809 8,809 129,975 129,975 129,975 2,551 2,551 48,697

1217 2216 424 1366 1114 62598 13729 1649 1533 2086

1219 2218 474 1491 848 19630 8207 1581 1559 2163

RTF

T-Flow T-DDPM T-Diff TPF T-Syn TVAE

629 1212 1219 751 2191 2219 239 45 467 612 48 1562 293 1097 50 1022 61865 63708 915 13594 15730 150 1609 146 1564 607 59 2181

-

1219 2217 411 1355 1579 1532 1834

1219 2218 434 1459 1535 1547 2168

932 2203 433 1016 2550 1462 1699

1049 81 324 924 83 84 1774

Table 4: Representative column-level cases for cardinality fidelity. Each generative-model entry is a distinct-value count. Entries for generative models are highlighted as First, Second, and Third within each row by closeness to the real statistics. High-Cardinality. We study cardinality because it determines whether the synthetic table preserves the state space needed by downstream analytical queries. For discrete columns, cardinality measures whether the generated table retains the distinct values present in the real data. This is especially important for highcardinality attributes, where a column may contain hundreds, thousands, or even more distinct states. If a generative model drops a large portion of these values, then later filtering, grouping, and subgroup queries no longer operate over the same support as the real table. In this case, the synthetic table may appear reasonable under aggregate distributional metrics while still failing to represent the set of values that downstream analyses depend on. Figure 5 shows that this failure is common on high-cardinality discrete columns. Each point represents one generative-model– column pair, with the x-axis showing the number of distinct values in the real column and the y-axis showing the number of distinct values generated synthetically. The diagonal corresponds to perfect cardinality preservation. Many points fall far below this line, especially when the real column contains thousands to tens of thousands of distinct values. This means that the synthetic table often preserves only a fraction of the real discrete support. The problem is also generative-model-dependent: some generative models remain close to the diagonal for several columns, while others collapse high-cardinality support to only hundreds of synthetic values. Table 4 shows the support loss at the raw column level. For columns with relatively moderate cardinality, some generative models can preserve the distinct count almost exactly. For example, on dataset c14, CTGAN, T-Flow, T-Diff, and TPF recover nearly all distinct values for columns with around 1K–12K real states. However, the table also shows that this behavior is far from universal. Several generative models collapse the same columns to only a small fraction of the real support: BayesNet repeatedly generates only 64 or 256 distinct values, ForestDiffusion often reduces thousands of real states to hundreds, and TVAE can collapse high-cardinality columns even more severely. The failure becomes especially clear on the largest-support columns. For dataset c18, the real designation column has 33,068 distinct values, but BayesNet generates only 255, ForestDiffusion only 74, and CTGAN only 9,679. Similarly, the title column has 96,779 real distinct values, while several generative models produce far fewer synthetic states. This support collapse directly affects query-centric fidelity: queries that group by these attributes, filter on specific values, or compare rare category behavior may lose many real states

entirely. Therefore, high-cardinality support preservation remains a major weakness of current synthetic tabular generative models. Figure 10 makes this query object concrete. It counts support by title on dataset c18, so any synthetic collapse in the distinct support of title directly changes which groups exist and how often they appear. Finding 3: Current SOTA generative models preserve global conditional structure more reliably than local-slice conditional structure (Figure 6). Conditional Failure in Local Slices. A local-slice query evaluates the same group-by aggregation as its global counterpart, but only within a filtered subset of the table. Figure 10 shows one grounded local-slice example on dataset c6. The paired global counterpart is obtained by deleting only the predicate WHERE "Subtopic" = ’Linear Transformations’, while keeping the SELECT, GROUP BY, and ORDER BY clauses unchanged. Therefore, the local query isolates whether the same two-dimensional group structure is preserved specifically within the “Linear Transformations” subpopulation, rather than across the full table. Figure 6 shows that local-slice conditional queries are usually less faithfully preserved than their global counterparts. Each point represents one generative model on one dataset under one paired conditional query. The x-axis reports the fidelity score of the global counterpart, and the y-axis reports the score of the matching localslice query. The points below the diagonal correspond to cases where the same query structure becomes less faithful after restricting the analysis to a subpopulation. This pattern holds across most paired comparisons. Among the 322 generative-model–dataset–query points, 82.6% lie below the diagonal, meaning that the local-slice query has lower fidelity than its global counterpart. The average drop is 0.11. Thus, current generative models often preserve whole-table conditional structure better than the same structure inside filtered subpopulations. This result shows that high global conditional fidelity is not sufficient: a synthetic table may answer broad conditional queries accurately while still failing on the local slices that analysts use to study specific subgroups. Finding 4: Current SOTA generative models fail to preserve rare cases and tail distributions, which leads to poor fidelity on tail-focused analytical queries (Figure 7 and Figure 8)

9

1 L1 2D count L2 Median slice L3 Group quantile L4 Time-bucket

0.6

0.65 Score

Local-slice score

0.8

Overall Coverage Size

0.6

0.4

0.55 0.2 10%

0

0

0.2

0.4 0.6 0.8 Global counterpart score

8%

6%

4%

3%

2%

1% 0.5% 0.1%

Tail threshold

1

Figure 7: Tail overall, coverage, and size under progressively rarer support. All three absolute scores decline as the rarity threshold tightens.

Figure 6: Local-slice vs. global-counterpart conditional scores. Each paired global query G1–G4 is obtained from the matching local-slice SQL by deleting only the single WHERE predicate; all other query structure is unchanged. Points below the diagonal indicate lower fidelity after restricting the same conditional query to a filtered local slice. The four local templates correspond to C-L1 to C-L4; see Appendix C.1, Table 10 for the full catalog.

Figure 8 shows that this trend is not caused by a single weak generative model. Each curve tracks one generative model’s tail overall score relative to its own 10% baseline, and most generative models decline as the threshold becomes more extreme. This means that the rare-region failure is broadly shared across current SOTA generative models: even when a generative model performs reasonably at a coarse tail threshold, its fidelity usually deteriorates as the evaluation moves to rarer support. Together, the two figures show that preserving rare cases and tail distributions remains a difficult regime for synthetic tabular generation. We next evaluate whether these rare-region failures affect the actual tail-query templates in TabQueryBench. These templates are ordinary analytical SQL queries: they use a tail predicate as part of a larger aggregation. Figure 10 shows one representative grounded tail query. It filters dataset m8 to the top 3% of balance and then groups the surviving rows by previous, so fidelity depends on preserving both the tail predicate and the post-filter support pattern. The benchmark results show that current generative models perform poorly on these tail-focused queries. Figure 8 reports the mean tail-query score for each generative model as the tail threshold becomes more restrictive. Even at the loose 10% threshold, most generative models are far below strong query-centric fidelity: only RealTabFormer exceeds 0.60, while many generative models are already below 0.40, and several are near or below 0.25. As the threshold tightens, the scores generally decline further. In the low-support ultra-tail region, RealTabFormer remains the strongest generative model but still stays below 0.50, while most other generative models remain substantially lower. This confirms that the rare-region failures observed in Figures 7–8 propagate to actual analytical queries: when a query depends on rare cases, current synthetic tables often cannot provide reliable answers. Missingness. Missingness provides a useful contrast. Many generative models match the overall amount of missingness well, but the benchmark still evaluates structured missingness through explicit query objects rather than raw null counts alone. Figure 10 shows one representative grounded missingness query: it groups dataset c5 by odor and asks whether the missing rate of stalk-root is preserved within each discrete state. A generative model can therefore match the global missing fraction while still failing this

Tail and Rarity. We first ask whether current generative models preserve the rare regions of individual columns, before evaluating full tail queries. Given a tail threshold 𝜏, such as 1%, we define the tail region of a column from the real data. For a continuous column, the tail region is defined by the empirical CDF, using the lower and/or upper quantile ranges whose total probability mass is 𝜏. For a categorical column, we sort categories by frequency in ascending order and define the tail region as the least frequent categories whose cumulative frequency reaches 𝜏. This definition lets us test whether a generative model preserves rare numerical ranges and rare categorical states, independently of downstream query. We measure this tail preservation using two complementary scores. The first is tail size, which asks whether the synthetic table assigns the right amount of mass to the real tail. We select the tail region on the real data, count how many synthetic rows fall into that real-defined region, and normalize by the corresponding number of real rows. The second is tail coverage, which asks whether the synthetic table recovers the same tail region. We compute the tail region separately on the real and synthetic data, then measure their overlap normalized by the size of the real tail region. For categorical columns, this overlap is the number of shared tail categories; for continuous columns, it is the length or measure of the shared tail interval. The tail overall score averages tail size and tail coverage, so it captures both the amount of synthetic mass placed in the tail and the identity of the rare values or ranges being preserved. Figures 7 and 8 show that current generative models have limited ability to preserve rare regions, and this limitation becomes more severe as the tail threshold tightens. Figure 7 reports the absolute tail scores averaged across generative models. Even at the relatively loose 10% threshold, the tail overall score is only around 0.64, with tail coverage around 0.55 and tail size around 0.69. As the threshold decreases toward 0.1%, all three scores decline, indicating that generative models preserve both less tail mass and less tail identity when the benchmark focuses on rarer cases. 10

Overall

Subgroup

Conditional

Tail / Rarity

Missingness

Cardinality / Range

Category Query overall ↑ Structure ↑ Categorical Numerical Mix

0.68±0.07 0.59±0.06 0.63±0.06

0.77±0.06 0.52±0.08 0.55±0.05

Size ↑

Global Structure ↑ Local Slice ↑ Tail Coverage ↑ Tail Size ↑ Marginal ↑ Broad Co-Missingness ↑ Discrete ↑ Continuous ↑

0.78±0.06 0.54±0.10 0.63±0.05

0.64±0.06 0.32±0.06 0.51±0.06

0.63±0.10 0.15±0.09 0.56±0.09

0.14±0.12 0.07±0.07 0.10±0.12

0.31±0.24 0.60±0.11 0.23±0.18

0.93±0.01 0.99±0.01 0.94±0.02

0.93±0.01 0.99±0.01 0.94±0.02

0.66±0.21 0.58±0.25 0.85±0.14

0.88±0.16 0.92±0.08 0.94±0.08

Table 5: Query taxonomy by data regime. Rows are ranked within each column and highlighted as First, Second, and Third.

0.65

RealTabFormer

RTF

0.5

BayesNet

Mean query score

Mean tail overall score

0.6

Bayes

0.4

ARF CTGAN

0.3

TPF TVAE

T-Flow T-DDPM

0.2

ARF

TabDiff TabSyn TabbyFlow ForestDiff.

0.6

TabDDPM

0.55

Higher quality

T-Syn

CTGAN

T-Diff

0.1 10%

F-Diff

8%

6%

4%

2%

1%

%

0.5

0.1

0.5

Tail threshold

1

5

10

50

100

200

Mean total runtime (min, log scale)

Figure 8: Tail degradation across individual generative models. Each curve tracks one generative model’s mean tail overall score as the rarity threshold tightens.

Figure 9: Query-centric fidelity and runtime define a Pareto trade-off on 9 datasets(C2, C7, C14, M4, M6, M8, N3, N6, N11). achieve mid-to-high query-centric fidelity, but require substantially more runtime than BayesNet and ARF for no aggregate fidelity gain. CTGAN, TVAE, and TabDDPM are relatively inexpensive, but they fall into a lower-fidelity regime. Overall, the figure sharpens the cost finding that would be missed by looking only at the best score: the highest-fidelity generative model is not the most cost-effective one, and BayesNet offers the strongest practical quality-cost balance on the common-9 runtime slice.

query if the alignment between missingness and the conditioning attribute is distorted. 5.3.2 Breaking Down by Categorical vs. Numerical vs. Mixed. We further separate results across categorical, numerical, and mixed datasets. This analysis tests whether the same query family fails differently depending on the schema regime, rather than treating all tabular datasets as one homogeneous benchmark pool. Table 5 shows that queries on categorical datasets are often about preserving discrete support, subgroup identity, and rare-state membership, whereas queries on numerical datasets more often test range structure, continuous conditional behavior, and tail sizes.

5.4

Higher cost

TVAE

0.5

%

TabPFGen

5.5

The Stability of Query Generation Pipeline

Because TabQueryBench uses a grounded query-generation pipeline, the stability of the benchmark instrument itself matters. The relevant question is whether rerunning the LLM-assisted SQL grounding step changes the comparative conclusions among generative models. We therefore rerun the same query-generation pipeline three times on a common probe set of nine datasets, selected to cover different column types and dataset sizes. Each run uses the same Stage 1 template library and the same dataset schemas; the only regenerated object is the dataset-specific SQL query set produced by the grounding and realization procedure. To quantify stability, we report three complementary rankingbased statistics. For each dataset–generative-model pair, let 𝑟 1, 𝑟 2, 𝑟 3 denote the generative-model ranks under the three regenerated query sets, and define the local rank spread as Δ𝑟 = max(𝑟 1, 𝑟 2, 𝑟 3 ) − min(𝑟 1, 𝑟 2, 𝑟 3 ). We then compute Kendall’s 𝑊 on each dataset to measure agreement of the available generative-model ranking across the three runs, where 𝑊 = 1 indicates perfect concordance. Finally, we compute Spearman’s 𝜌 between each pair of run-level ranking vectors and report the mean pairwise correlation. Table 6 summarizes the results. Across the 92 non-missing dataset– generative-model cells, 83.7% of cells move by at most one rank. At the dataset level, the mean Kendall’s 𝑊 is 0.927, and the mean pairwise Spearman’s 𝜌 is 0.903 across the three regeneration pairs.

Fidelity-cost Pareto tradeoff.

Finding 5: There is a clear cost-fidelity tradeoff in tabular data generation and BayesNet achieves the best balance (Table 9). Figure 9 shows that query-centric fidelity and runtime separate into distinct practical regimes once we mark the Pareto frontier. RealTabFormer achieves the highest mean Query overall score on the common-9 datasets, but it is also by far the most expensive generative model, with a mean total runtime close to 150 minutes. It therefore occupies a high-fidelity/high-cost regime rather than a universally attractive operating point. BayesNet, by contrast, lies on the low-cost/strong-fidelity end of the frontier: it is orders of magnitude cheaper than the highest-cost neural generative models while still delivering one of the strongest query scores. ARF remains a cost-effective mid-cost alternative relative to the most expensive generative models, but it does not lie on the strict frontier because BayesNet is both cheaper and slightly stronger. The remaining generative models occupy less favorable regions of the tradeoff space. TabDiff, TabPFGen, TabSyn, and TabbyFlow 11

Summary statistic

Value

Δ𝑟 = 0 Δ𝑟 = 1 Δ𝑟 = 2–3

55.4% 28.3% 16.3%

Mean Kendall’s 𝑊 Mean pairwise Spearman’s 𝜌

0.927 0.903

synthetic data should preserve the analytical operations that downstream users are likely to run. This extension would be especially important for analytical questions which cross table boundaries. Privacy vs. fidelity. Query-centric fidelity also has a natural connection to privacy evaluation, where synthetic-data releases are often judged through privacy-utility tradeoffs, differential-privacy mechanisms, or threat-specific leakage analyses [26, 46, 61, 64]. The same structural assessors that measure whether a synthetic table preserves analytical behavior can reveal how much rare support, local subgroup structure, or high-cardinality information survives generation. Query-centric fidelity is not a formal privacy metric, and it is not a substitute for membership-inference, attribute-inference, or differential-privacy analyses. Instead, it highlights a fidelityprivacy tension that users must interpret for their specific settings. In particular, privacy risk is often contextual: some releases may tolerate accurate aggregates, while others may treat specific query outputs as sensitive information. When query outputs are the objects users want to protect, query-centric evaluation provides a useful diagnostic because the same query structure can be used to measure how much protected signal leaks through the synthetic table. A synthetic table that preserves rare values or high-cardinality keys may be more useful for analytics, but such preservation should be read together with privacy reports and threat models. LLM-assisted SQL grounding. TabQueryBench uses an LLM only inside the constrained template-to-SQL realization step: the profiling, binding, and validation are deterministic, and the validation rejects invalid SQL, incompatible realizations, and queries that fail execution. Our stability study suggests that the resulting conclusions are broadly stable across regenerated query sets, but LLMassisted SQL grounding remains a source of benchmark-instrument variance. The current study reruns the same grounding pipeline, but it does not yet compare multiple SQL realization models or multiple prompting policies. Future work could audit SQL grounding across different LLMs, compare family-level and model-level rankings under these alternative realizations, and include additional human or rule-based checks for ambiguous schema bindings.

Table 6: Ranking stability across three regenerated query sets on a 9-dataset probe set. The upper block reports the fractions of dataset–generative-model cells by local rank spread Δ𝑟 , and the lower block reports the rank-agreement statistics.

The few visibly unstable cases are concentrated in m4 and n3; outside these outliers, regenerated query sets largely preserve the overall comparative ranking and usually keep the top-performing generative model unchanged.

6

DISCUSSION AND LIMITATIONS

Query-centric fidelity and distance-based fidelity. Query-centric fidelity is not intended to replace distance-based fidelity. The two views measure different aspects of synthetic data quality. Distancebased metrics summarize global resemblance over columns or distributions, while query-centric fidelity asks whether the synthetic table preserves the answers to analytical operations that downstream users are likely to run. Our results show why both views are needed. This makes query-centric fidelity a complementary evaluation layer rather than a substitute for statistical similarity. In practice, the right evaluation depends on the downstream use case. Users who need mock data for coarse schema exploration may care more about global resemblance, while users who need data system testing or analytics on synthetic data may need stronger query-centric fidelity. Toward query-aware synthetic-data generation. TabQueryBench also suggests a model-development direction: future tabular generative models can use query families as training, validation, or model-selection signals. The family-level breakdown identifies where current models lose analytical structure, rather than only reporting an aggregate query score. For example, high-cardinality results point to the need for better discrete support preservation. Tail and rarity results point to the need for better rare-region generation. Local conditional results point to the need for models that preserve filtered subpopulation structure. These targets are concrete because they correspond to executable SQL queries and measurable query answers. A query-aware generative model could therefore optimize not only global distributional resemblance, but also take these analytical queries as part of the loss function or reward signals so that the generated table can support these operations. Multi-table query-centric evaluation. The current benchmark is designed to support the single-table setting, where reusable analytical query patterns can be grounded across heterogeneous schemas through schema-aware realization. A natural next step is multi-table evaluation. In that setting, synthetic data must preserve not only column distributions and single-table query answers, but also joins, foreign-key structure, join selectivity, fanout behavior, and query answers across related tables. Multi-table query-centric fidelity would extend the same principle used in TabQueryBench:

7

CONCLUSIONS

TabQueryBench reframes synthetic tabular data evaluation around analytical query answers. Instead of asking only whether a synthetic table matches the real table under aggregate statistical distances, it asks whether the table preserves the answers to queries that downstream users are likely to run. We instantiate this idea with 49 datasets and 11 generative models. The benchmark uses reusable query templates, schema-aware grounding, and five query families. Our evaluation shows that distance-based fidelity does not reliably predict query-centric fidelity. These results suggest that synthetic tabular data should be evaluated by the analytical operations it can support. Aggregate resemblance alone is not enough. Future work should study how query-centric fidelity predicts downstream utility in concrete applications, such as dashboard accuracy, textto-SQL evaluation, and model development on synthetic data. Another important direction is multi-table evaluation, where synthetic data must preserve joins, foreign-key structure, and query answers across related tables. We believe that TabQueryBench can serve as a foundation for future synthetic-data benchmarks that treat the preservation of analytical query answers as a first-class objective. 12

Appendix

B

OVERALL BENCHMARK SUMMARY AND QUERY SCORE HEATMAP

Dataset ARF Bayes CTGAN F-Diff RTF T-Flow T-DDPM T-Diff TPF T-Syn TVAE

A

TEMPLATE SOURCES, TAXONOMY, AND GROUNDING

c2 c3 c4 c5 c6 c7 c8 c9 c10 c11 c12 c13 c14 c15 c16 c17 c18 c19 c20 m1 m2 m4 m5 m6 m7 m8 m9 m10 m11 m12 n1 n2 n3 n4 n5 n6 n7 n8 n9 n10 n11 n12 n14 n15 n16 n17 n18 n19 n20

This appendix records the template-side sources, taxonomy, grounding rules, and the compact classical distance baseline used in the current benchmark release. The main text gives the high-level pipeline; here we keep only the details needed to audit where templates came from, how the released 44-template inventory is organized, how dataset-specific realizations are constrained, and how the classical reference metrics are defined.

A.1

Template Sources

The current appendix-facing template inventory surfaces 44 templates in total. Table 7 lists the public benchmark, repository, documentation, and paper sources that directly contribute to one or more released templates.

Direct source

Year / venue

Templates

ClickBench [11] H2O db-benchmark [23] TPC-H / TPC-DS benchmark specs [65, 66] TPC-DS benchmark family [65] TPC-H qgen [66] TPC-DS qualification repo [12] TPC-DS Altinity repo [1] RTABench order_events slice [63] BigQuery approximate aggregate docs [19] ClickHouse aggregate docs [10] Snowflake PERCENTILE_CONT docs [58] Snowflake WIDTH_BUCKET docs [59] Trino aggregate docs [67] Apache Druid SQL-function docs [2] Preserving Missing Data Distribution in Synthetic Data [70]

2022 | benchmark 2020 | benchmark 1999–2017 | benchmark specs 2017 | benchmark spec 1999 | benchmark Public | TPC-DS family Public | TPC-DS family 2025 | benchmark + GitHub Current | docs Current | docs Current | docs Current | docs Current | docs Current | docs 2023 | paper

8 9 1 3 4 2 2 2 2 1 2 1 3 1 3

Table 7: Sources that directly contribute to one or more templates in the released library.

A.2

0.885 0.700 0.986 0.677 0.726 0.886 0.778 0.495 0.654 0.909 0.739 0.571 0.733 0.707 0.677 0.669 0.589 0.883 0.618 0.691 0.601 0.775 0.444 0.924 0.802 0.626 0.739 0.689 0.598 0.406 0.804 0.829 0.522 0.481 0.512 0.821 0.749 0.437 0.446 0.833 0.828 0.537 0.421 0.623 0.699 0.383 0.371 0.456 0.486

0.889 0.677 0.969 0.622 0.699 0.870 0.989 0.526 0.653 0.889 0.814 0.754 0.822 0.693 0.593 0.629 0.438 0.654 0.969 0.764 0.628 0.920 0.443 0.756 0.916 0.741 0.655 0.900 0.729 0.449 0.811 0.930 0.523 0.498 0.646 0.794 0.900 0.459 0.500 0.844 0.844 0.536 0.439 0.816 0.837 0.638 0.353 0.580 0.490

0.860 0.691 0.875 0.591 0.547 0.832 0.926 0.482 0.843 0.792 0.749 0.579 0.683 0.752 0.629 0.752 0.549 0.553 0.846 0.759 0.574 0.723 0.343 0.476 0.749 0.496 0.708 0.749 0.616 0.371 0.739 0.687 0.348 0.381 0.538 0.656 0.679 0.328 0.490 0.823 0.833 0.766 0.399 0.356 0.806 0.537 0.303 0.660 0.350

0.655 0.765 0.749 0.597 0.615 0.768 0.759 0.404 0.653 0.631 0.722 0.852 0.713 0.630 0.542 0.545 0.450 0.693 0.595 0.730 0.575 0.737 0.406 0.650 0.726 0.514 0.586 0.646 0.560 0.342 0.783 0.489 0.495 0.626 0.473 0.756 0.718 0.483 0.375 0.803 0.802 0.398 0.500 0.706 0.927 0.353 0.405 0.348 0.400

0.951 0.588 0.913 0.676 0.721 0.874 0.989 0.961 0.889 0.978 0.843 0.883 0.793 0.762 0.721 0.485 0.758 0.463 0.959 0.623 0.542 0.887 0.580 0.869 0.878 0.671 0.759 0.866 0.868 0.677 0.780 0.886 0.661 0.606 0.620 0.817 0.799 0.519 0.581 0.840 0.846 0.908 0.428 0.765 0.931 0.617 0.502 0.625 0.658

0.617 0.573 0.771 0.551 0.585 0.738 0.615 0.581 0.659 0.655 0.125 0.857 0.912 0.690 0.381 0.420 0.387 0.623 0.594 0.640 0.580 0.775 0.456 0.703 0.786 0.553 0.512 0.702 0.719 0.435 0.662 0.555 0.485 0.515 0.532 0.814 0.712 0.433 0.447 0.832 0.838 0.472 TF 0.623 0.927 TF 0.420 0.347 0.400

0.647 0.520 0.773 0.537 0.640 0.765 0.612 0.376 0.653 0.655 TF 0.604 0.516 TF TF 0.262 TF TF 0.491 0.594 0.512 0.651 0.346 0.608 0.727 0.530 0.561 0.550 0.613 TF 0.653 0.508 0.505 0.406 0.332 0.683 0.660 TF 0.450 0.842 0.844 0.478 0.426 0.378 0.691 0.518 TF 0.309 0.322

0.636 0.572 0.759 0.546 0.582 0.748 0.631 0.593 0.981 0.977 TF 0.841 0.915 0.696 0.528 0.562 TF 0.694 TF 0.649 TF 0.779 0.444 0.715 0.782 0.530 0.595 0.834 0.680 0.406 0.926 0.594 0.462 0.637 0.488 0.930 0.788 0.485 0.333 0.926 0.925 0.365 0.526 0.716 0.927 0.365 0.408 0.406 TF

0.796 0.678 0.849 0.756 0.679 0.821 0.772 0.457 0.625 0.993 0.739 0.444 0.962 0.768 TF 0.651 TF 0.701 0.614 0.696 0.626 0.775 0.452 0.686 0.742 0.559 0.658 0.696 0.664 0.405 0.844 0.703 0.478 0.503 0.520 0.895 0.743 0.108 0.446 0.835 0.843 0.465 0.419 0.612 0.927 0.385 0.404 0.479 0.496

0.705 0.478 0.756 0.536 0.588 0.736 0.661 0.569 0.093 0.651 0.861 0.857 0.665 0.684 0.516 0.407 0.428 0.683 0.587 0.641 0.546 0.784 0.461 0.707 0.786 0.527 0.558 0.698 0.697 0.412 0.801 0.575 0.539 0.303 0.283 0.723 0.504 0.417 0.341 0.357 0.736 0.422 0.454 0.688 0.803 0.346 0.418 0.413 0.465

0.833 0.593 0.809 0.570 0.623 0.858 0.744 0.470 0.738 0.903 0.556 0.585 0.579 0.408 0.419 0.615 0.541 0.571 0.817 0.582 0.582 0.688 0.332 0.440 0.595 0.488 0.546 0.761 0.599 0.359 0.776 0.684 0.412 0.398 0.519 0.655 0.738 0.339 0.481 0.834 0.831 0.805 0.407 0.589 0.787 0.524 0.317 0.530 0.343

Table 9: Dataset-by-generative-model overall query score heatmap. TF indicates technical failure due to unsupported generation or prohibitively long runtime.

Classical Distance-Based Fidelity

This appendix records a compact classical baseline. We use Jensen–Shannon distance (JSD) for aligned categorical or discretized distributions, Kolmogorov– Smirnov distance (KS) for the largest cumulative-distribution gap in continuous variables, Total Variation distance (TVD) for discrete support discrepancy, and Wasserstein distance for continuous numerical shift, normalized so that results remain comparable across columns. Distance-based Fidelity

Query Fidelity

Cost (min)

Generator Dist. overall ↑ REAL ARF BayesNet CTGAN ForestDiffusion RealTabFormer TabbyFlow TabDDPM TabDiff TabPFGen TabSyn TVAE

1.00±0.00 0.91±0.13 0.86±0.16 0.81±0.13 0.66±0.19 0.91±0.15 0.77±0.27 0.64±0.24 0.84±0.25 0.86±0.14 0.68±0.26 0.80±0.18

JSD ↓

KS ↓

TVD ↓

0.00±0.00 0.00±0.00 0.00±0.00 0.17±0.28 0.06±0.05 0.15±0.27 0.19±0.28 0.16±0.16 0.17±0.27 0.22±0.23 0.27±0.17 0.20±0.22 0.60±0.37 0.15±0.13 0.59±0.37 0.15±0.24 0.05±0.09 0.14±0.23 0.36±0.36 0.09±0.06 0.34±0.37 0.53±0.36 0.29±0.27 0.51±0.36 0.21±0.27 0.06±0.06 0.19±0.26 0.23±0.26 0.10±0.09 0.21±0.26 0.61±0.42 0.07±0.08 0.60±0.43 0.26±0.28 0.25±0.17 0.24±0.28

Wasserstein ↓ 0.00±0.00 0.02±0.02 0.05±0.07 0.08±0.07 0.05±0.07 0.01±0.01 0.04±0.04 0.19±0.18 0.03±0.05 0.02±0.02 0.02±0.03 0.08±0.08

Query overall ↑ Subgroup ↑ Conditional ↑ Tail / Rarity ↑ Missingness ↑ Cardinality / Range ↑ Train ↓ Gen. ↓ 1.00±0.00 0.66±0.16 0.70±0.17 0.63±0.17 0.61±0.15 0.75±0.15 0.60±0.17 0.55±0.14 0.66±0.19 0.65±0.18 0.57±0.17 0.60±0.16

1.00±0.00 0.67±0.37 0.68±0.35 0.63±0.36 0.61±0.37 0.76±0.30 0.59±0.36 0.55±0.39 0.63±0.37 0.63±0.39 0.57±0.37 0.61±0.34

1.00±0.00 0.50±0.35 0.52±0.32 0.47±0.31 0.41±0.33 0.64±0.27 0.44±0.32 0.43±0.34 0.46±0.32 0.48±0.33 0.43±0.30 0.43±0.29

1.00±0.00 0.35±0.26 0.43±0.30 0.24±0.19 0.08±0.14 0.45±0.27 0.19±0.24 0.13±0.21 0.09±0.19 0.27±0.23 0.18±0.27 0.21±0.18

1.00±0.00 0.97±0.14 0.95±0.15 0.96±0.14 0.95±0.15 0.97±0.12 0.94±0.16 0.96±0.15 0.95±0.16 0.96±0.15 0.94±0.16 0.94±0.16

1.00±0.00 0.79±0.26 0.90±0.18 0.80±0.20 0.58±0.31 0.89±0.17 0.76±0.33 0.61±0.31 0.85±0.30 0.84±0.23 0.65±0.33 0.75±0.21

— 4.85 0.24 3.74 11.2 144.0 14.0 0.84 50.2 0.003 41.6 4.54

— 0.70 0.15 0.77 0.22 3.58 5.59 0.69 11.7 39.9 0.27 0.23

Table 8: Overall benchmark summary across 11 synthetic tabular generative models averaged over 49 datasets. ↑ means higheris-better and ↓ means lower-is-better. REAL is the self-comparison reference row, and top generative models are highlighted as First, Second, and Third. The last two columns report mean training and generation time in minutes.

13

C

TEMPLATE, DATASET, AND GENERATIVE MODEL CATALOGS

This appendix collects the long-form catalogs used by the released benchmark: the 44-template taxonomy inventory, the 49-dataset roster, and the active generative-model roster. These tables are grouped here to preserve space while keeping the released benchmark assets in one auditable location.

C.1

44-Template Taxonomy Catalog

The compact taxonomy summary in Appendix A collapses related templates into family-level rows. The longtable below expands that same appendix-facing taxonomy back to the template level while keeping all 44 released templates in one consistent row format. Table 10: 44-template taxonomy catalog. Template ID

No.

Canonical template

Family

Subfamily

Upstream source

grp_cond_rate grp_ratio_2cond ds_within_grp_share win_part_avg 2d_tgt_rate binned_num_grp_avg grp_disp_rank ds_base_gated_rank flt_2d_grp_count med_flt_num cond_grp_quants rta_time_bucket_cnt rta_time_bucket_mavg

C-G1 C-G2 C-G3 C-G4 C-G5 C-G6 C-G7 C-G8 C-L1 C-L2 C-L3 C-L4 C-L5

Condition Rate by Group Two-Condition Ratio by Group Item Share Within Each Group Partition Average by Group 2D Target-Rate Surface Binned Group Average Group Dispersion Rank Baseline-Gated Extreme Ranking Filtered 2D Count Surface Filtered Median Slice Filtered Group Quantiles Filtered Time-Bucket Count Time-Bucket Moving Average by Group

Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional Conditional

Global Structure Global Structure Global Structure Global Structure Global Structure Global Structure Global Structure Global Structure Local Slice Local Slice Local Slice Local Slice Local Slice

TPC-H, TPC-DS TPC-DS TPC-DS qualification H2O db-benchmark TPC-DS Snowflake WIDTH_BUCKET docs Trino aggregate docs TPC-DS Altinity TPC-DS Snowflake PERCENTILE_CONT docs ClickHouse aggregate docs RTABench order_events RTABench order_events

cb_grp_distinct_topk cb_flt_distinct_topk h2o_grp_sum grp_avg_num guarded_grp_avg ds_topk_grp_sum h2o_2d_grp_sum 2d_grp_avg cb_grp_summary_topk tpch_2d_summary h2o_2d_robust tpch_max_agg_win wtd_topk_sum cb_grp_count cb_flt_grp_count cb_2d_topk_count

S-1.1 S-1.2 S-1.3 S-1.4 S-1.5 S-1.6 S-1.7 S-1.8 S-1.9 S-1.10 S-1.11 S-1.12 S-1.13 S-2.1 S-2.2 S-2.3

Distinct-Coverage Ranking by Group Filtered Distinct-Coverage Ranking by Group Total-Measure Ranking by Group Mean-Measure Comparison by Group Support-Gated Mean Comparison by Group Filtered Total-Measure Ranking by Group Two-Dimensional Total-Measure Ranking Two-Dimensional Mean Comparison Support-and-Mean Ranking by Group Filtered Two-Dimensional Measure Summary Two-Dimensional Robust Measure Summary Top Aggregate Winner by Group Support-Gated Weighted Total Ranking by Group Group Count Distribution Filtered Group Count Ranking Two-Dimensional Group Count Ranking

Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup Subgroup

Structure Structure Structure Structure Structure Structure Structure Structure Structure Structure Structure Structure Structure Size Size Size

ClickBench ClickBench H2O db-benchmark H2O db-benchmark H2O db-benchmark TPC-DS qualification H2O db-benchmark H2O db-benchmark ClickBench TPC-H qgen H2O db-benchmark TPC-H qgen BigQuery approx docs ClickBench ClickBench ClickBench

quant_tail_slice global_z_outliers h2o_topn_in_grp ds_subgrp_base_outlier tpch_rel_total_thr tpch_thr_grp_rank grp_pct_point thr_rarity_cdf

T-1.1 T-1.2 T-1.3 T-1.4 T-2.1 T-2.2 T-2.3 T-2.4

Top-Quantile Tail Members Global Z-Score Outliers Within-Group Top-N Extremes Subgroup-Relative Outliers Groups Above a Total-Share Threshold Above-Threshold Group Ranking Group-wise Percentile Level Threshold Exceedance Rarity

Tail / Rarity Tail / Rarity Tail / Rarity Tail / Rarity Tail / Rarity Tail / Rarity Tail / Rarity Tail / Rarity

Tail Coverage Tail Coverage Tail Coverage Tail Coverage Tail Size Tail Size Tail Size Tail Size

Snowflake PERCENTILE_CONT docs Trino aggregate docs H2O db-benchmark TPC-DS Altinity TPC-H qgen TPC-H qgen BigQuery approx docs Apache Druid SQL docs

miss_rate_marg miss_rate_disc_state miss_rate_cont_bucket

M-1.1 M-2.1 M-2.2

Marginal Missing Rate Missing Rate by Discrete State Missing Rate by Continuous Bucket

Missingness Missingness Missingness

Marginal Missingness Broad Co-Missingness Broad Co-Missingness

Preserving Missing Data Distribution in Synthetic Data Preserving Missing Data Distribution in Synthetic Data Preserving Missing Data Distribution in Synthetic Data

card_supp_rank_prof card_dist_share_prof card_cont_range_env card_hi_card_resp

K-1.1 K-1.2 K-1.3 K-2.1

Support Rank Profile Distinct Share Profile Continuous Range Envelope Profile High-Cardinality Response Stability

Cardinality / Range Cardinality / Range Cardinality / Range Cardinality / Range

Discrete Discrete Continuous Discrete

ClickBench ClickBench Trino aggregate docs H2O db-benchmark

C.2

Dataset Catalog Table 11: 49 dataset catalog.

Series No.

Original dataset name

C C C C C C C C C C C

Jungle Chess 2pcs Raw Endgame Complete Car Evaluation Splice junction Gene Sequences Chess King Rook Vs King Pawn Mushroom Dataset For Assessing ML In Higher Education Nursery Phishing Websites Amazon Employee Access Challenge Poker Hand Connect 4

c1 c2 c3 c4 c5 c6 c7 c8 c9 c10 c11

Rows

Cols

44819 1728 3189 3196 8416 9546 12960 11055 32769 1025010 67557

7 7 3 37 23 8 9 31 10 11 43

Original source URL https://www.openml.org/d/41027 https://archive.ics.uci.edu/dataset/19/car+evaluation https://archive.ics.uci.edu/dataset/69/molecular+biology+splice+junction+gene+sequences https://archive.ics.uci.edu/dataset/22/chess+king+rook+vs+king+pawn https://archive.ics.uci.edu/dataset/73/mushroom https://archive.ics.uci.edu/dataset/1031/dataset+for+assessing+mathematics+learning+in+higher+education https://www.openml.org/d/26 https://archive.ics.uci.edu/dataset/327/phishing+websites https://www.kaggle.com/c/amazon-employee-access-challenge https://archive.ics.uci.edu/dataset/158/poker+hand https://archive.ics.uci.edu/dataset/26/connect-4 Continued on next page 14

Table 11: 49 dataset catalog (continued). Series No.

Original dataset name

C C C C C C C C M M M M M M M M M M M N N N N N N N N N N N N N N N N N N N

Internet-Advertisements Us Census Data 1990 Cat In The Dat Cat In The Dat Ii Fivethirtyeight Comic Characters Dataset Netflix Shows Wine Reviews Trending YouTube Video Statistics and Comments Remote Worker Productivity Car Insurance Claim Prediction Hotel Booking Demand Medical Insurance Charges Predict Students Dropout And Academic Success Online Shoppers Purchasing Intention Dataset Stroke Prediction Dataset Bank Marketing HR Analytics: Job Change of Data Scientists Mobile Price Classification Health Insurance Cross Sell Prediction Spambase Airfoil Self-Noise Wine Quality Communities And Crime Superconductivity BEED: Bangalore EEG Epilepsy Dataset Anuran Calls Mfccs Secom Pen Based Recognition Of Handwritten Digits Dry Bean Dataset Magic Gamma Telescope Skin Segmentation Combined Cycle Power Plant First-Order Theorem Proving COIL2000 Credit Card Fraud Detection Statlog Shuttle APS Failure at Scania Trucks Fashion-MNIST

c12 c13 c14 c15 c16 c17 c18 c19 m1 m2 m12 m4 m5 m6 m7 m8 m9 m10 m11 n1 n2 n3 n4 n5 n6 n7 n8 n9 n10 n11 n12 n13 n14 n15 n16 n17 n18 n19

Rows

Cols

Original source URL

3279 2458285 300000 600000 6896 8807 129971 40949 1500 58592 119390 2772 4424 12330 5110 45211 19158 2000 381109 4601 1503 4898 1994 21263 8000 7195 1567 10992 13611 19019 245057 9568 2000 9822 284807 14500 76000 70000

1559 69 25 25 13 12 14 16 30 44 32 7 37 18 12 17 14 21 12 58 6 12 128 82 17 26 593 17 17 11 4 5 52 86 31 10 171 785

https://www.openml.org/d/40978 https://archive.ics.uci.edu/dataset/116/us+census+data+1990 https://www.kaggle.com/c/cat-in-the-dat https://www.kaggle.com/c/cat-in-the-dat-ii https://www.kaggle.com/datasets/fivethirtyeight/fivethirtyeight-comic-characters-dataset https://www.kaggle.com/datasets/shivamb/netflix-shows https://www.kaggle.com/datasets/zynicide/wine-reviews https://www.kaggle.com/datasets/datasnaek/youtube-new https://huggingface.co/datasets/nprak26/remote-worker-productivity https://www.kaggle.com/datasets/ifteshanajnin/carinsuranceclaimprediction-classification https://www.kaggle.com/datasets/jessemostipak/hotel-booking-demand https://huggingface.co/datasets/rahulvyasm/medical_insurance_data https://archive.ics.uci.edu/dataset/697/predict+students+dropout+and+academic+success https://archive.ics.uci.edu/dataset/468/online+shoppers+purchasing+intention+dataset https://www.kaggle.com/datasets/fedesoriano/stroke-prediction-dataset https://archive.ics.uci.edu/dataset/222/bank+marketing https://www.kaggle.com/datasets/arashnic/hr-analytics-job-change-of-data-scientists https://www.kaggle.com/datasets/iabhishekofficial/mobile-price-classification https://www.kaggle.com/datasets/anmolkumar/health-insurance-cross-sell-prediction https://www.openml.org/d/44 https://archive.ics.uci.edu/dataset/291/airfoil+self-noise https://archive.ics.uci.edu/dataset/186/wine+quality https://archive.ics.uci.edu/dataset/183/communities+and+crime https://www.openml.org/d/44964 https://archive.ics.uci.edu/dataset/1134/beed:+bangalore+eeg+epilepsy+dataset https://archive.ics.uci.edu/dataset/406/anuran+calls+mfccs https://archive.ics.uci.edu/dataset/179/secom https://archive.ics.uci.edu/dataset/81/pen-based+recognition+of+handwritten+digits https://archive.ics.uci.edu/dataset/602/dry+bean+dataset https://archive.ics.uci.edu/dataset/159/magic+gamma+telescope https://archive.ics.uci.edu/dataset/229/skin+segmentation https://archive.ics.uci.edu/dataset/294/combined+cycle+power+plant https://www.openml.org/d/44663 https://www.openml.org/d/298 https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud https://archive.ics.uci.edu/dataset/148/statlog+shuttle https://archive.ics.uci.edu/dataset/421/aps+failure+at+scania+trucks https://www.openml.org/d/40996

15

REFERENCES

[23] H2O.ai. 2020. Database-like Operations Benchmark. https://h2oai.github.io/dbbenchmark/. https://h2oai.github.io/db-benchmark/ Benchmark report and reproducibility repository. [24] Lasse Hansen, Nabeel Seedat, Mihaela van der Schaar, and Andrija Petrovic. 2023. Reimagining Synthetic Tabular Data Generation through Data-Centric AI: A Comprehensive Benchmark. In Advances in Neural Information Processing Systems, Vol. 36. https://papers.nips.cc/paper_files/paper/2023/hash/ 6aa9a05b929fb08ff46a58cab6cf860d-Abstract-Datasets_and_Benchmarks.html Datasets and Benchmarks Track. [25] Mikel Hernandez, Gorka Epelde, Ane Alberdi, Rodrigo Cilla, and Debbie Rankin. 2023. Synthetic Tabular Data Evaluation in the Health Domain Covering Resemblance, Utility, and Privacy Dimensions. Methods of Information in Medicine 62, S 01 (2023), e19–e38. https://doi.org/10.1055/s-0042-1760247 [26] Julian Höllig and Michaela Geierhos. 2025. Utility Meets Privacy: A Critical Evaluation of Tabular Data Synthesizers. IEEE Access 13 (2025), 44497–44509. https://doi.org/10.1109/ACCESS.2025.3549680 [27] Valter Hudovernik, Martin Jurković, and Erik Štrumbelj. 2024. Benchmarking the Fidelity and Utility of Synthetic Relational Data. arXiv preprint arXiv:2410.03411 (2024). https://doi.org/10.48550/arXiv.2410.03411 [28] Hugging Face. [n.d.]. Hugging Face Datasets. https://huggingface.co/datasets. https://huggingface.co/datasets Accessed: 2026-07-01. [29] Tobias Hyrup, Anton Danholt Lautrup, Arthur Zimek, and Peter Schneider-Kamp. 2024. Sharing is CAIRing: Characterizing Principles and Assessing Properties of Universal Privacy Evaluation for Synthetic Tabular Data. Machine Learning with Applications 18 (2024), 100608. https://doi.org/10.1016/j.mlwa.2024.100608 [30] Xi Jiang, Shinan Liu, Aaron Gember-Jacobson, Arjun Nitin Bhagoji, Paul Schmitt, Francesco Bronzino, and Nick Feamster. 2024. Netdiffusion: Network data augmentation through protocol-constrained traffic generation. Proceedings of the ACM on Measurement and Analysis of Computing Systems 8, 1 (2024), 1–32. [31] Xi Jiang, Shinan Liu, Aaron Gember-Jacobson, Paul Schmitt, Francesco Bronzino, and Nick Feamster. 2023. Generative, high-fidelity network traces. In Proceedings of the 22nd ACM Workshop on Hot Topics in Networks. 131–138. [32] Xiangjian Jiang, Nikola Simidjievski, and Mateja Jamnik. 2026. TabStruct: Measuring Structural Fidelity of Tabular Data. In The Fourteenth International Conference on Learning Representations. https://openreview.net/forum?id=XOPH34Extq Oral presentation. [33] Alexia Jolicoeur-Martineau, Kilian Fatras, and Tal Kachman. 2024. Generating and Imputing Tabular Data via Diffusion and Flow-based Gradient-Boosted Trees. In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research), Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li (Eds.), Vol. 238. PMLR, 1288–1296. https://proceedings. mlr.press/v238/jolicoeur-martineau24a.html [34] Bayrem Kaabachi, Jérémie Despraz, Thierry Meurers, Karen Otte, Mehmed Halilovic, Bogdan Kulynych, Fabian Prasser, and Jean Louis Raisaro. 2025. A Scoping Review of Privacy and Utility Metrics in Medical Synthetic Data. npj Digital Medicine 8, 1 (2025), 60. https://doi.org/10.1038/s41746-024-01359-3 [35] Kaggle. [n.d.]. Kaggle. https://www.kaggle.com. https://www.kaggle.com Accessed: 2026-07-01. [36] Shivasankari Kannan, Yeounoh Chung, Amita Gondi, Tristan Swadell, and Fatma Ozcan. 2025. High-Fidelity And Complex Test Data Generation For Google SQL Code Generation Services. https://doi.org/10.48550/arXiv.2504.17203 arXiv:2504.17203 [cs.DB] [37] Markelle Kelly, Rachel Longjohn, and Kolby Nottingham. [n.d.]. The UCI Machine Learning Repository. https://archive.ics.uci.edu. https://archive.ics.uci.edu Accessed: 2026-07-01. [38] Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. 2023. TabDDPM: Modelling Tabular Data with Diffusion Models. In Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research), Vol. 202. 17564–17579. https://proceedings.mlr.press/v202/ kotelnikov23a.html [39] Anton D Lautrup, Tobias Hyrup, Arthur Zimek, and Peter Schneider-Kamp. 2025. Syntheval: a framework for detailed utility and privacy evaluation of tabular synthetic data. Data Mining and Knowledge Discovery 39, 1 (2025), 6. https://doi.org/10.1007/s10618-024-01081-4 [40] Junwei Ma, Apoorv Dankar, George Stein, Guangwei Yu, and Anthony Caterini. 2023. TabPFGen – Tabular Data Generation with TabPFN. In NeurIPS 2023 Second Table Representation Learning Workshop. https://openreview.net/forum?id= 4MkkNsAEmO [41] Mark Mazumder, Colby Banbury, Xiaozhe Yao, Bojan Karlaš, William Gaviria Rojas, Sudnya Diamos, Greg Diamos, Lynn He, Alicia Parrish, Hannah Rose Kirk, Jessica Quaye, Charvi Rastogi, Douwe Kiela, David Jurado, David Kanter, Rafael Mosquera, Will Cukierski, Juan Ciro, Lora Aroyo, Bilge Acun, Lingjiao Chen, Mehul Raje, Max Bartolo, Evan Sabri Eyuboglu, Amirata Ghorbani, Emmett Goodman, Addison Howard, Oana Inel, Tariq Kane, Christine R. Kirkpatrick, D. Sculley, Tzu-Sheng Kuo, Jonas W. Mueller, Tristan Thrush, Joaquin Vanschoren, Margaret Warren, Adina Williams, Serena Yeung, Newsha Ardalani, Praveen Paritosh, Ce Zhang, James Y. Zou, Carole-Jean Wu, Cody Coleman, Andrew Y.

[1] Altinity. [n.d.]. ClickHouse TPC-DS Benchmark. https://github.com/Altinity/tpcds. https://github.com/Altinity/tpc-ds [2] Apache Druid. [n.d.]. All Druid SQL Functions. https://druid.apache.org/docs/ latest/querying/sql-functions/ [3] Patricia A. Apellániz, Ana Jiménez, Borja Arroyo Galende, Juan Parras, and Santiago Zazo. 2024. Synthetic Tabular Data Validation: A Divergence-Based Approach. IEEE Access 12 (2024), 103895–103907. https://doi.org/10.1109/ACCESS. 2024.3434582 [4] arka420. [n.d.]. Insurance-Cost-Project-Using-SQL. https://github.com/arka420/ Insurance-Cost-Project-Using-SQL. https://github.com/arka420/InsuranceCost-Project-Using-SQL [5] Andrew G. Bates, Iva Špakulová, Iain Dove, and Andrew Mealor. 2019. ONS Methodology Working Paper Series Number 16: Synthetic Data Pilot. Technical Report. Office for National Statistics. https://www.ons.gov.uk/methodology/ methodologicalpublications/generalmethodology/onsworkingpaperseries/ onsmethodologyworkingpaperseriesnumber16syntheticdatapilot Published 15 January 2019. [6] Bernd Bischl, Giuseppe Casalicchio, Matthias Feurer, Pieter Gijsbers, Frank Hutter, Michel Lang, Rafael G. Mantovani, Jan N. van Rijn, and Joaquin Vanschoren. 2021. OpenML Benchmarking Suites. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, Vol. 1. [7] Hasan Alp Caferoğlu, Mehmet Serhat Çelik, and Özgür Ulusoy. 2025. SING-SQL: A Synthetic Data Generation Framework for In-Domain Text-to-SQL Translation. https://doi.org/10.48550/arXiv.2509.25672 arXiv:2509.25672 [cs.AI] [8] Andrew Chu, Xi Jiang, Shinan Liu, Arjun Bhagoji, Francesco Bronzino, Paul Schmitt, and Nick Feamster. 2024. Feasibility of state space models for network traffic generation. In Proceedings of the 2024 SIGCOMM Workshop on Networks for AI Computing. 9–17. [9] Andrew Chu, Xi Jiang, Shinan Liu, Arjun Bhagoji, Francesco Bronzino, Paul Schmitt, and Nick Feamster. 2026. Netssm: Multi-flow and state-aware network trace generation using state-space models. Proceedings of the ACM on Networking 4, CoNEXT1 (2026), 1–24. [10] ClickHouse. [n.d.]. Aggregate Functions. https://clickhouse.com/docs/sqlreference/aggregate-functions/reference [11] ClickHouse. 2022. ClickBench: A Benchmark for Analytical Databases. https://github.com/ClickHouse/ClickBench. https://github.com/ClickHouse/ ClickBench GitHub benchmark. [12] CWI Database Architectures Group. [n.d.]. Reproducing TPC-DS Qualification Results. https://github.com/cwida/tpcds-result-reproduction. https://github. com/cwida/tpcds-result-reproduction [13] DataCebo. 2026. Welcome to SDGym! https://docs.sdv.dev/sdgym Official documentation. [14] Bailu Ding, Surajit Chaudhuri, Johannes Gehrke, and Vivek R. Narasayya. 2021. DSB: A Decision Support Benchmark for Workload-Driven and Traditional Database Systems. Proceedings of the VLDB Endowment 14, 13 (2021), 3376–3388. https://doi.org/10.14778/3484224.3484234 [15] Junhui Ding, Xinchen Zhang, Xiaohui Xie, and Shinan Liu. 2026. TraceCodec: A Compiler-Backed Neural Codec for Stateful Multi-Flow Network Traffic Traces. arXiv preprint arXiv:2605.29941 (2026). [16] Yuntao Du and Ninghui Li. 2025. Systematic Assessment of Tabular Data Synthesis. In Proceedings of the 2025 ACM SIGSAC Conference on Computer and Communications Security. 2414–2428. https://doi.org/10.1145/3719027.3765067 [17] Khaled El Emam, Lucy Mosquera, Xi Fang, and Alaa El-Hussuna. 2022. Utility Metrics for Evaluating Synthetic Health Data Generation Methods: Validation Study. JMIR Medical Informatics 10, 4 (2022), e35734. https://doi.org/10.2196/ 35734 [18] Nick Erickson, Lennart Purucker, Andrej Tschalzev, David Holzmüller, Prateek Mutalik Desai, David Salinas, and Frank Hutter. 2026. TabArena: A Living Benchmark for Machine Learning on Tabular Data. In The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track. https://openreview.net/forum?id=jZqCqpCLdU [19] Google Cloud. [n.d.]. GoogleSQL Approximate Aggregate Functions. https://cloud.google.com/bigquery/docs/reference/standardsql/approximate_aggregate_functions [20] Léo Grinsztajn, Edouard Oyallon, and Gaël Varoquaux. 2022. Why Do Tree-Based Models Still Outperform Deep Learning on Typical Tabular Data?. In Advances in Neural Information Processing Systems, Vol. 35. 507–520. [21] Ragini Gupta, Shinan Liu, Ruixiao Zhang, Xinyue Hu, Xiaoyang Wang, Hadjer Benkraouda, Pranav Kommaraju, Phuong Cao, Nick Feamster, and Klara Nahrstedt. 2025. Generative active adaptation for drifting and imbalanced network intrusion detection. arXiv preprint arXiv:2503.03022 (2025). [22] Andrés Guzmán-Cordero, Floor Eijkelboom, and Jan-Willem van de Meent. 2025. Exponential Family Variational Flow Matching for Tabular Data Generation. In Forty-second International Conference on Machine Learning. https://openreview. net/forum?id=kjtvCSkSsy

16

Ng, Peter Mattson, and Vijay Janapa Reddi. 2023. DataPerf: Benchmarks for Data-Centric AI Development. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, Vol. 3. https://papers.nips. cc/paper_files/paper/2023/hash/112db88215e25b3ae2750e9eefcded94-AbstractDatasets_and_Benchmarks.html [42] Raghunath Othayoth Nambiar and Meikel Poess. 2006. The Making of TPC-DS. In Proceedings of the 32nd International Conference on Very Large Data Bases. VLDB Endowment, 1049–1058. [43] Noshin Nawar Nawar. [n.d.]. Exploratory Analysis of Car Evaluation Dataset with SQL. https://github.com/nehanawar025/Exploratory-Analysis-of-CarEvaluation-Dataset-with-SQL. https://github.com/nehanawar025/ExploratoryAnalysis-of-Car-Evaluation-Dataset-with-SQL [44] OpenML. [n.d.]. OpenML. https://www.openml.org. https://www.openml.org Accessed: 2026-07-01. [45] Neha Patki, Roy Wedge, and Kalyan Veeramachaneni. 2016. The Synthetic Data Vault. In 2016 IEEE International Conference on Data Science and Advanced Analytics. 399–410. https://doi.org/10.1109/DSAA.2016.49 [46] Personal Data Protection Commission Singapore. 2024. Proposed Guide on Synthetic Data Generation. Guidance document. https: //www.pdpc.gov.sg/organisations/resources/guidance-by-topic/proposedguide-on-synthetic-data-generation Published 24 September 2024. [47] Haoyue Ping, Julia Stoyanovich, and Bill Howe. 2017. DataSynthesizer: PrivacyPreserving Synthetic Datasets. In Proceedings of the 29th International Conference on Scientific and Statistical Database Management (SSDBM ’17). 42:1–42:5. https: //doi.org/10.1145/3085504.3091117 [48] Meikel Poess and Chris Floyd. 2000. New TPC Benchmarks for Decision Support and Web Commerce. ACM SIGMOD Record 29, 4 (2000), 64–71. https://doi.org/ 10.1145/369275.369291 [49] Meikel Poess, Raghunath Othayoth Nambiar, and David Walrath. 2007. Why You Should Run TPC-DS: A Workload Analysis. In Proceedings of the 33rd International Conference on Very Large Data Bases. 1138–1149. https://www.vldb.org/conf/ 2007/papers/industrial/p1138-poess.pdf [50] Meikel Poess, Bryan Smith, Lubor Kollár, and Per-Åke Larson. 2002. TPC-DS, Taking Decision Support Benchmarking to the Next Level. In Proceedings of the 2002 ACM SIGMOD International Conference on Management of Data. 582–587. https://doi.org/10.1145/564691.564759 Syn[51] Zhaozhi Qian, Rob Davis, and Mihaela van der Schaar. 2023. thcity: A Benchmark Framework for Diverse Use Cases of Tabular Synthetic Data. In Advances in Neural Information Processing Systems, Vol. 36. https://papers.nips.cc/paper_files/paper/2023/hash/ 09723c9f291f6056fd1885081859c186-Abstract-Datasets_and_Benchmarks.html Datasets and Benchmarks Track. [52] Gabriele Santangelo, Giovanna Nicora, Riccardo Bellazzi, and Arianna Dagliati. 2025. How Good Is Your Synthetic Data? SynthRO, a Dashboard to Evaluate and Benchmark Synthetic Tabular Data. BMC Medical Informatics and Decision Making 25, 1 (2025), 89. https://doi.org/10.1186/s12911-024-02731-9 [53] Aniruddha Sen, Christine Task, Dhruv Kapur, Gary Howarth, and Karan Bhagat. 2023. Diverse Community Data for Benchmarking Data Privacy Algorithms. Advances in Neural Information Processing Systems 36 (2023), 51409–51420. [54] Shagufta-DataAnalyst. [n.d.]. insurance-sql-analysis. https://github.com/ Shagufta-DataAnalyst/insurance-sql-analysis. https://github.com/ShaguftaDataAnalyst/insurance-sql-analysis [55] Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec. 2025. TabDiff: a Mixed-type Diffusion Model for Tabular Data Generation. In The Thirteenth International Conference on Learning Representations. https://openreview.net/forum?id=swvURjrt8z [56] Andrey Sidorenko, Michael Platzer, Mario Scriminaci, and Paul Tiwald. 2025. Benchmarking Synthetic Tabular Data: A Multi-Dimensional Evaluation Framework. arXiv preprint arXiv:2504.01908 (2025). https://doi.org/10.48550/arXiv. 2504.01908 [57] Joshua Snoke, Gillian M. Raab, Beata Nowok, Chris Dibben, and Aleksandra Slavković. 2018. General and Specific Utility Measures for Synthetic Data. Journal of the Royal Statistical Society Series A: Statistics in Society 181, 3 (2018), 663–688. https://doi.org/10.1111/rssa.12358 [58] Snowflake Inc. 2026. PERCENTILE_CONT — Snowflake Documentation. https: //docs.snowflake.com/en/sql-reference/functions/percentile_cont. https://docs. snowflake.com/en/sql-reference/functions/percentile_cont [59] Snowflake Inc. 2026. WIDTH_BUCKET — Snowflake Documentation. https: //docs.snowflake.com/en/sql-reference/functions/width_bucket. https://docs. snowflake.com/en/sql-reference/functions/width_bucket [60] Aivin V. Solatorio and Olivier Dupriez. 2023. REaLTabFormer: Generating Realistic Relational and Tabular Data using Transformers. arXiv preprint arXiv:2302.02041 (2023). https://doi.org/10.48550/arXiv.2302.02041 [61] Yuchao Tao, Ryan McKenna, Michael Hay, Ashwin Machanavajjhala, and Gerome Miklau. 2021. Benchmarking Differentially Private Synthetic Data Generation Algorithms. arXiv preprint arXiv:2112.09238 (2021). https://doi.org/10.48550/ arXiv.2112.09238

[62] Christine Task, Karan Bhagat, and Gary Howarth. 2023. SDNist v2: Deidentified Data Report Tool. https://doi.org/10.18434/MDS2-2943 [63] Timescale. 2025. RTABench: A Benchmark for Real-Time Analytics Applications. https://github.com/timescale/rtabench. https://github.com/timescale/rtabench GitHub benchmark. [64] Van Tran, Shinan Liu, Tian Li, and Nick Feamster. 2025. Quantifying the Privacy Implications of High-Fidelity Synthetic Network Traffic. arXiv preprint arXiv:2511.20497 (2025). [65] Transaction Processing Performance Council. [n.d.]. TPC Benchmark DS (TPCDS) Standard Specification. https://www.tpc.org/tpcds/. https://www.tpc.org/ tpcds/ [66] Transaction Processing Performance Council. [n.d.]. TPC Benchmark H (TPC-H) Standard Specification. https://www.tpc.org/tpch/. https://www.tpc.org/tpch/ [67] Trino Software Foundation. 2026. Aggregate Functions — Trino Documentation. https://trino.io/docs/current/functions/aggregate.html. https://trino.io/docs/ current/functions/aggregate.html [68] Boris van Breugel, Zhaozhi Qian, and Mihaela van der Schaar. 2023. Synthetic Data, Real Errors: How (Not) to Publish and Use Synthetic Data. In Proceedings of the 40th International Conference on Machine Learning (Proceedings of Machine Learning Research), Vol. 202. 34793–34808. https://proceedings.mlr.press/v202/ van-breugel23a.html [69] Jason Walonoski, Mark Kramer, Joseph Nichols, Andre Quina, Chris Moesel, Dylan Hall, Carlton Duffett, Kudakwashe Dube, Thomas Gallagher, and Scott McLachlan. 2018. Synthea: An Approach, Method, and Software Mechanism for Generating Synthetic Patients and the Synthetic Electronic Health Care Record. Journal of the American Medical Informatics Association 25, 3 (2018), 230–238. https://doi.org/10.1093/jamia/ocx079 [70] Xinyue Wang, Hafiz Asif, and Jaideep Vaidya. 2023. Preserving Missing Data Distribution in Synthetic Data. In Proceedings of the ACM Web Conference 2023. 2110–2121. https://doi.org/10.1145/3543507.3583297 [71] David S. Watson, Kristin Blesch, Jan Kapar, and Marvin N. Wright. 2023. Adversarial Random Forests for Density Estimation and Generative Modeling. In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research), Vol. 206. 5357–5375. https://proceedings.mlr.press/v206/watson23a.html [72] Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. 2019. Modeling Tabular Data using Conditional GAN. In Advances in Neural Information Processing Systems, Vol. 32. https://proceedings.neurips.cc/paper/ 2019/hash/254ed7d2de3b23ab10936522dd547b78-Abstract.html [73] Zhangchen Xu, Fengqing Jiang, Luyao Niu, Yuntian Deng, Radha Poovendran, Yejin Choi, and Bill Yuchen Lin. 2025. Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing. In The Thirteenth International Conference on Learning Representations. [74] Hengrui Zhang, Jiani Zhang, Zhengyuan Shen, Balasubramaniam Srinivasan, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2024. MixedType Tabular Data Synthesis with Score-Based Diffusion in Latent Space. In The Twelfth International Conference on Learning Representations. https://openreview. net/forum?id=4Ay23yeuz0 [75] Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean-Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. 2021. DatasetGAN: Efficient Labeled Data Factory with Minimal Human Effort. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 10145–10155.

17

Related documents

• TrajectoryDB: A New Database for Agent Trajectories#673620
• Preserving contextual information in cultural heritage metadata through multidimensional knowledge graphs#673621
• X-DigCheck: Co-Evolving Application Profiles and Knowledge Graphs, Demonstrated on the RTI Documentation of Rupe Magna#673622
• Open Tabular Insight Extraction: Where Do We Stand, and Where Should We Go?#673623
• Protocol effects on feature-based hardware-Trojan detection across Trust-Hub families#673624