Infodat: Benchmark & Data Integration

• Infodat is an umbrella term for benchmark datasets and systems that convert raw textual and statistical data into structured, queryable infographic metadata.
• The Infodat benchmark enables text-to-complex infographic generation by mapping narrative input to detailed metadata for multi-sub-chart visuals.
• Its construction pipeline combines automated synthesis with human verification to ensure accurate extraction of chart types, axes, statistics, and layouts.

Searching arXiv for papers mentioning “Infodat” and closely related uses. Infodat denotes two closely related research constructs. In its most explicit current usage, it is the benchmark dataset introduced for text-to-complex statistical infographic metadata generation, where each sample links a textual document to structured metadata describing a multi-sub-chart statistical infographic (Ghosh et al., 26 Jul 2025). In a broader interpretive usage across several other works, “Infodat” functions as a label for systems that make data integrated, queryable, transparent, accessible, or computationally actionable, including open-data warehouses, personal-data transparency interfaces, cognitively accessible visualization frameworks, legal-compliance ontologies, information-theoretic dependency analyzers, city-scale multimodal datasets, and federated scientific metadata search systems (Santore et al., 2019, Schufrin et al., 2020, Wu et al., 2023, Leyva-Sánchez et al., 27 Mar 2026, Hu et al., 29 May 2026, Cappi et al., 6 May 2026, 0907.5538).

1. Definition and semantic scope

Infodat, as formally introduced in the infographic-generation literature, is a benchmark for mapping an input text document $T$ to infographic metadata $M$, with downstream code generation represented as $C = g(f(T))$ (Ghosh et al., 26 Jul 2025). The target metadata includes a global title and summary together with sub-chart-level descriptors such as chart kind, axes, statistics, local text, relative positions, dimensions, fonts, background, alignment, and local summaries. The benchmark is specifically aimed at complex statistical infographics, defined as statistical visuals composed of more than one sub-chart, each with its own axes and titles (Ghosh et al., 26 Jul 2025).

A broader reading emerges from adjacent literature. BIOD is described as a centralized, integrated repository of Brazilian public open data with server-side querying and pre-aggregation (Santore et al., 2019). TransparencyVis is explicitly framed as an “Infodat” tool because it converts raw GDPR exports into unified and comparable visual overviews of personal data (Schufrin et al., 2020). Work on people with Intellectual and Developmental Disabilities reinterprets “Infodat” as information and data accessibility grounded in lived experience and cognitively accessible visualization (Wu et al., 2023). DAOnt turns Data Act provisions into RDF/OWL representations and SPARQL-checkable obligations, permissions, and prohibitions (Leyva-Sánchez et al., 27 Mar 2026). InfoAtlas treats Infodat as real-time dependency analysis through zero-shot mutual information estimation (Hu et al., 29 May 2026). The Padua traffic resource and the XML-based planetary search system similarly instantiate information-rich, multi-source layers over raw observational data (Cappi et al., 6 May 2026, 0907.5538).

A plausible implication is that “Infodat” has evolved into an umbrella term for an intermediate information layer between raw data and downstream human or machine use: a layer that standardizes schema, exposes semantics, and enables inspection, reasoning, or generation.

2. Infodat as a benchmark dataset for statistical infographics

The benchmark version of Infodat was introduced alongside “Infogen” (Ghosh et al., 26 Jul 2025). Its purpose is to support text-to-complex statistical infographic metadata generation, a task for which no prior dataset existed in the form of text-only input paired with richly structured metadata for a multi-sub-chart statistical infographic (Ghosh et al., 26 Jul 2025). The benchmark is therefore positioned not as a chart-understanding or chart-summarization resource, but as a planning resource for complete infographic construction.

Each Infodat sample links a synthetic but realistic input document to metadata that is sufficient to reconstruct the infographic as code (Ghosh et al., 26 Jul 2025). The document is a single continuous passage that contains the narrative, statistics, and local descriptions needed for reconstruction, while deliberately excluding explicit structural hints such as chart types, the number of sub-charts, or terms such as “section,” “chart,” or “image” (Ghosh et al., 26 Jul 2025). The metadata then decomposes the infographic into global and sub-chart components. At the global level, it contains a title and summary. At the local level, each subchart_k includes kind, axis, stats, text, position_chart, position_chart_text, background, dimensions, fonts, alignment, and summary (Ghosh et al., 26 Jul 2025).

The scale of the dataset is fixed by the reported corpus statistics. Infodat contains 3,463 document–metadata pairs, with an average input length of 185.46 words, median 162 words, average 15.07 sentences, average metadata length of 341.02 words, median 300 words, and average 14.54 sentences in the metadata (Ghosh et al., 26 Jul 2025). The metadata contains an average of 2.15 sub-charts, with median 2, minimum 1, and maximum 21 (Ghosh et al., 26 Jul 2025). The data split is 80 / 5 / 15 for train, validation, and test (Ghosh et al., 26 Jul 2025).

These properties distinguish Infodat from text-to-simple-chart settings. It requires simultaneous long-context textual understanding, numerical extraction, sub-chart segmentation, and layout planning. The chart coverage includes bar charts, line charts, pie charts, histograms, and stacked or grouped variants (Ghosh et al., 26 Jul 2025).

3. Construction pipeline and metadata schema

Infodat is built through a semi-automated pipeline grounded in the Pew dataset of statistical charts scraped from Pew Research Center articles (Ghosh et al., 26 Jul 2025). The first stage is selection. Because the original source includes many simple single-chart figures, GPT-4o is used with few-shot prompting to classify which images qualify as “complex infographics,” retaining 3,463 such images (Ghosh et al., 26 Jul 2025). The selection criterion is explicitly information density and the presence of multiple sub-charts with associated title, heading, and context.

The second stage is synthesis of the input document. GPT-4o with vision is instructed to produce a detailed textual description of each selected infographic under strong constraints: the output must be a single continuous passage; it must include the relevant statistics and associated text; and it must not mention chart type, number of sections, or source-image metadata such as logos or footnotes (Ghosh et al., 26 Jul 2025). This produces the benchmark input $T$.

The third stage is synthesis of the target metadata $M$. GPT-4o with chain-of-thought and few-shot examples is used to count sub-charts and generate detailed metadata for each one, including chart type, axes and units, data points, associated text, relative positions, dimensions, fonts, and background (Ghosh et al., 26 Jul 2025). This output is then human verified and corrected by graduate interns with experience in statistical infographics and familiarity with Pew-style statistics (Ghosh et al., 26 Jul 2025). The guidelines require checking the number of sub-charts, chart type, axis labels and units, statistical values, and relative positioning.

The resulting schema is not merely descriptive; it is operational. The metadata is designed so that a downstream model can convert it into executable code for Plotly or Plotnine (Ghosh et al., 26 Jul 2025). This design choice makes Infodat a planning benchmark rather than an image-captioning dataset. It also establishes a clean separation between semantic extraction from text and graphical realization.

4. Infodat within the Infogen framework and its evaluation regime

Infodat is the substrate for the first stage of Infogen, a two-stage framework in which fine-tuned LLMs generate metadata and a subsequent component converts metadata into infographic code (Ghosh et al., 26 Jul 2025). The first stage learns the mapping $M = f(T)$ by training on Infodat’s text–metadata pairs. The second stage applies a coder model to produce code $C = g(M)$, with a feedback model iterating over the code up to five times (Ghosh et al., 26 Jul 2025).

The reported training setup uses multiple fine-tuned LLM families, including LLAMA 3, Qwen2, and Phi-3, with QLoRA fine-tuning and DPO alignment (Ghosh et al., 26 Jul 2025). A ranker LLM then selects among multiple metadata candidates. This makes Infodat central not only as a benchmark but as the intermediate representation governing the full text-to-infographic pipeline.

Evaluation on Infodat includes both automatic and human protocols (Ghosh et al., 26 Jul 2025). The automatic metrics are designed for metadata generation rather than image similarity. They include Subchart Accuracy, Root Square Error (RSE) for sub-chart count, Subchart Type Accuracy, Statistical Accuracy, and ROUGE-L for title, global summary, and sub-chart summaries (Ghosh et al., 26 Jul 2025). Human evaluation is applied to final infographics generated from Infodat-based models on 35% of the test set and scores Readability, Visual Appeal, and Data Accuracy and Alignment on a 1–5 scale (Ghosh et al., 26 Jul 2025). In those results, GPT-4o (20-shot) scores 3.4, 2.8, and 2.4; Phi3 (DPO) scores 3.7, 3.2, and 3.4; and Infogen scores 4.1, 3.8, and 4.1 on the three dimensions, respectively (Ghosh et al., 26 Jul 2025).

The benchmark therefore formalizes a layered problem: identifying how many sub-charts the document implies, assigning the correct chart type to each, extracting the statistics, attaching the correct local insight text, and recovering a plausible layout. In the reported qualitative analysis, common failure modes for non-Infogen systems include mis-counting sub-charts, merging conceptually distinct sub-charts, and producing cluttered or misaligned layouts (Ghosh et al., 26 Jul 2025).

5. Broader “Infodat” architectures across research domains

Several other papers use “Infodat” as a conceptual frame rather than the name of a specific dataset. They consistently describe systems that insert an information-rich layer between heterogeneous data sources and downstream use.

Domain	Artifact	Infodat role
Open public data	BIOD	Integrated repository, query API, pre-aggregation
Personal privacy	TransparencyVis	Unified exploration of GDPR exports
Cognitive accessibility	IDD visualization study	Accessible, personal, narrative, shared data use
Legal compliance	DAOnt	RDF/OWL representation with SPARQL checking
Dependency analysis	InfoAtlas	Zero-shot mutual information estimation
Urban informatics	Padua traffic dataset	Multi-source spatio-temporal graph resource
Planetary science	IDIS XML search system	Federated metadata discovery across nodes

In BIOD, the relevant pattern is integration of fragmented public data into a single analytical back-end. BIOD integrates more than 300 GB of data across 24 tables, more than 900 columns, and nearly 3 billion rows, implemented in MonetDB and exposed through BlenDB with a metric/dimension query model and pre-aggregation (Santore et al., 2019). This places Infodat in the tradition of data warehouses and federated open-data hubs.

TransparencyVis addresses a different problem: legally available but cognitively inaccessible personal-data exports. It unifies heterogeneous GDPR exports from Google, Facebook, Instagram, and Twitter into a common model of File elements and Data elements, organized by a service-independent set of 10 categories, and visualized through a client-side interface with FileView, TimeView, and ListView (Schufrin et al., 2020). Here Infodat is information about one’s own data, designed to support perception, comprehension, and reflection.

For people with Intellectual and Developmental Disabilities, the relevant contribution is not data integration alone but cognitively accessible representation. The study is based on 15 semi-structured interviews with adults with self-identified mild–moderate IDD and their caregivers, and it argues that visualization should make data feel personal, function as a boundary object, support agency and self-advocacy, and work across modalities including physical, pictorial, and narrative forms (Wu et al., 2023). In this interpretation, Infodat denotes accessible participation in data-driven life rather than conventional dashboards.

DAOnt gives the term a legal-institutional meaning. It models the EU Data Act in OWL/RDF, reusing LKIF-Core, ODRL, and DPV, and operationalizes Article 4(1), 8(6), and 19(2)(a) for SPARQL-based compliance checking over obligations, permissions, and prohibitions (Leyva-Sánchez et al., 27 Mar 2026). This is an Infodat layer for data-sharing agreements: machine-readable information about who may do what with which data, under which legal conditions.

InfoAtlas reinterprets Infodat in information-theoretic terms. It treats mutual information estimation as a single-step inference problem, using a pretrained architecture that directly predicts MI in one forward pass, with reported $100\times$ speedup and generalization across dimensions, sample sizes, and real-world domains (Hu et al., 29 May 2026). The Padua traffic dataset and the preliminary XML-based planetary search system extend the same logic to domain infrastructures: the former integrates traffic volumes, transition matrices, travel times, flow residuals, POIs, demographics, weather, and road-network data through a spatio-temporal graph and Python class (Cappi et al., 6 May 2026); the latter implements an XML/XQuery/XPath-based federated metadata portal for planetary resources across distributed thematic nodes (0907.5538).

Taken together, these systems suggest a stable architectural pattern. “Infodat” repeatedly denotes a mediating representation: metadata instead of images, metric/dimension schemas instead of raw files, unified personal-data categories instead of ZIP archives, RDF/OWL norms instead of legal prose, or spatio-temporal graphs instead of isolated sensor tables.

6. Limitations, misconceptions, and future directions

The literature does not support the misconception that Infodat is a single standardized platform. Rather, the cited works suggest two layers of meaning: a proper noun for a benchmark dataset in infographic generation, and a broader interpretive label for infrastructures that make data more usable, comparable, or governable.

The benchmark dataset Infodat has explicit limitations. It is “relatively limited in size,” with 3,463 instances, and its domain coverage is narrow, being derived from English-language Pew Research content on social, political, and demographic statistics (Ghosh et al., 26 Jul 2025). The input documents are synthetically generated by GPT-4o from images, and the metadata is first auto-generated and then manually verified, so both sides of the benchmark inherit some dependence on the synthesis pipeline (Ghosh et al., 26 Jul 2025).

The broader “Infodat” systems also expose recurring constraints. BIOD requires custom treatment for each new source, faces data consistency issues across institutions, updates only annually, and uses a deliberately limited query language (Santore et al., 2019). TransparencyVis depends on fragile service-specific parsers, confronts clutter in high-volume datasets, and reported a first-iteration SUS of 65.4 (Schufrin et al., 2020). The IDD study is exploratory, with n = 15, self-reported diagnoses, and variation in caregiver participation (Wu et al., 2023). DAOnt currently covers 18 articles of the Data Act, with detailed operationalization for three, and some legal concepts remain semantically ambiguous in OWL/SPARQL form (Leyva-Sánchez et al., 27 Mar 2026). InfoAtlas relies on sliced MI above 20D and degrades for $n < 400$ (Hu et al., 29 May 2026). The Padua traffic dataset includes a data gap from 2026-03-19 to 2026-03-26, removes malfunction-prone sensors, and releases only aggregated trajectory statistics for privacy reasons (Cappi et al., 6 May 2026).

A plausible implication is that future work on Infodat will continue to expand along three axes. The first is coverage: larger datasets, more domains, more services, and more regulatory or scientific subfields. The second is formalization: richer intermediate schemas, controlled vocabularies, or ontologies that preserve semantics across heterogeneous sources. The third is actionability: more systems in which the intermediate layer is not only descriptive but operational, enabling code generation, compliance checking, server-side analytics, or real-time dependency analysis. Across its present uses, Infodat is best understood not as raw data itself, but as the structured informational interface through which raw data becomes integrated, interpretable, and usable.