Misviz-synth: Synthetic Chart Corpus

• Misviz-synth is a synthetic dataset of 81,814 charts designed to support the detection of misleading design features in visualizations.
• The dataset employs a two-phase construction pipeline that pairs realistic tabular data with controlled chart rendering and detailed metadata.
• Its applications span training multimodal models for chart de-rendering and misleader classification, though it is limited to standard chart types.

Misviz-synth is a fully synthetic corpus of misleading and non-misleading charts introduced to support the training and evaluation of models that detect “misleaders,” defined as design choices that distort the viewer’s interpretation of data. It accompanies the real-world Misviz benchmark and was designed to address the absence of large, diverse, and openly available training resources for this task. The dataset contains 81,814 charts, encodes 11 of the 12 misleader types used in Misviz, and provides perfectly paired visuals, data tables, and metadata for bar, line, and pie charts (Tonglet et al., 29 Aug 2025).

1. Definition and scope

Misviz-synth is a large synthetic dataset whose stated purpose is to support models that identify whether a chart is misleading and, when applicable, which specific misleader it contains (Tonglet et al., 29 Aug 2025). Its scope is defined by several fixed properties: 81,814 total charts; 11 encoded misleader types; three chart types—bar, line, and pie; and a train/validation/test split of 93\%/2\%/5\% (Tonglet et al., 29 Aug 2025). The omitted category is “discretized continuous variable,” which applies only to maps and is therefore excluded from the synthetic chart setting.

The class balance is intentionally asymmetric. Approximately 78\% of the charts are misleading, while approximately 22\% are clean (Tonglet et al., 29 Aug 2025). Each misleading chart carries exactly one misleader. This constraint produces a controlled supervision regime in which the label semantics are unambiguous at the instance level. A plausible implication is that Misviz-synth is especially suitable for supervised classification and controlled ablation studies, while being less representative of real-world cases in which multiple misleading devices may co-occur.

The dataset distribution across chart types is also explicit: bar charts account for 37,627 instances (46.0\%), line charts for 38,454 (47.0\%), and pie charts for 5,733 (7.0\%) (Tonglet et al., 29 Aug 2025). This concentration on standard chart forms reflects the dataset’s role as a tractable synthetic benchmark rather than an exhaustive ontology of visualization genres.

2. Construction pipeline

Misviz-synth is generated in two phases: table preparation and chart rendering (Tonglet et al., 29 Aug 2025). The first phase begins by harvesting real tables $T$ from Wikipedia (TabFact) and OurWorldInData. For each table, each column $C$ is assigned a primary type $$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$ together with zero or more secondary tags such as “Evenly spaced temporal” and “Is-part-of-whole” (Tonglet et al., 29 Aug 2025). Valid $(\text{key}, \text{value})$ column pairs or triples are then enumerated subject to three constraints: at least one Numerical column $v$, a natural key $k$ composed of Categorical and/or Temporal columns, and, when $k$ has more than one column, all but one key column are frozen to a random constant row value.

For each valid combination, a compatible chart type is selected from $\{\text{bar}, \text{line}, \text{pie}\}$, and a chart title is generated from templates such as “{value_col} by {key_col} in {year}” (Tonglet et al., 29 Aug 2025). This design couples real tabular content with synthetic rendering, preserving plausible data semantics while allowing systematic perturbation of chart design.

In the second phase, the system determines which misleader types are applicable given the prepared table and chart type. With probability $p_{\text{clean}}$, one clean chart is generated; otherwise, for each applicable misleader $m$, exactly one chart with misleader $C$0 is rendered (Tonglet et al., 29 Aug 2025). Rendering is performed by a hand-crafted Matplotlib function of the form plot_{chart_type, m}(P, θ), where $C$1 denotes misleader-specific parameters drawn from preset distributions. Random style variations are then applied, including font size, background color, and grid lines. For bar and line charts, axis metadata are additionally extracted into a four-field table with Seq, Axis, Label, and RelPos (Tonglet et al., 29 Aug 2025).

The accompanying pseudocode makes the generative logic explicit:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$0

This pipeline separates semantic preparation from visual realization. This suggests a modular benchmark design in which data semantics, chart compatibility, misleader applicability, and stylistic variation can be controlled independently.

3. Encoded misleaders and parameterization

Misviz-synth encodes 11 of the 12 misleader types used in the real-world Misviz benchmark (Tonglet et al., 29 Aug 2025). Among misleading charts—approximately 63,872 instances—the most frequent categories are inconsistent tick intervals with 15,653 charts (24.5\%), inverted axis with 12,013 (18.8\%), misrepresentation with 5,430 (8.5\%), truncated axis with 1,277 (2.0\%), and inappropriate use of pie chart with 1,467 (2.3\%) (Tonglet et al., 29 Aug 2025). The remaining categories—dual axis, binning, line misuse, item order, axis range, and 3D—each occupy between 1.3\% and 4.5\%.

Several misleaders are defined through explicit parameter distributions. For truncated-axis charts, letting $C$2, the pipeline draws $C$3 and sets $C$4 (Tonglet et al., 29 Aug 2025). For misrepresentation via nonlinear encoding, bar heights are set according to

$C$5

with $C$6 to induce under-scaling or over-scaling (Tonglet et al., 29 Aug 2025). For inconsistent tick intervals, ticks $C$7 are generated using step sizes drawn from two Gaussians $C$8 and $C$9 and then interleaved (Tonglet et al., 29 Aug 2025).

These parameterizations matter because the misleaders are not merely symbolic labels; they are instantiated through concrete rendering procedures. A plausible implication is that the dataset can support both label prediction and mechanism-sensitive learning, since the visual evidence for a given label is algorithmically coupled to a known transformation.

4. Data representation and file structure

Each dataset instance is stored as a JSON object together with an image in PNG or SVG format (Tonglet et al., 29 Aug 2025). The JSON schema contains the following fields: image_id, chart_type, misleader, title, data_table, axis_metadata for bar and line charts, and style_params (Tonglet et al., 29 Aug 2025). The data_table is represented as an array of records in which each record contains key field(s) and a value. The axis_metadata field is an array of objects with seq, axis, label, and rel_pos. The style_params field stores randomization choices such as font_size, grid, and color_theme.

The provided example illustrates the structure:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$1

This representation is notable because it links three modalities: raster or vector image content, structured table data, and extracted axis metadata. The paper explicitly identifies this multimodal organization as useful for training image + table + axis metadata models (Tonglet et al., 29 Aug 2025). This suggests that Misviz-synth is not only a vision benchmark but also a benchmark for cross-modal alignment between rendered charts and their underlying tabular semantics.

5. Rendering implementation and exemplar code

The rendering layer is implemented with hand-crafted Matplotlib functions (Tonglet et al., 29 Aug 2025). The summary includes several representative code fragments that clarify how specific misleaders are operationalized.

A truncated y-axis can be generated as follows:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$2

Misrepresentation through nonlinear bar height is illustrated by:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$3

Inconsistent tick intervals are produced by alternating step sizes:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$4

A dual y-axis example is also given:

$$C.type \in \{\text{Numerical}, \text{Categorical}, \text{Temporal}\}$$5

These snippets show that the synthetic process is not abstract annotation layered on preexisting images; it is direct procedural chart synthesis. This matters because the causal path from table content to rendered artifact is preserved, which in turn supports perfectly labeled training data and reproducible perturbation regimes.

6. Uses, limitations, and interpretive boundaries

The stated applications of Misviz-synth fall into three categories: training multimodal models to classify or localize misleaders; fine-tuning and evaluating chart de-renderers such as DePlot on perfectly labeled synthetic data; and serving as a controlled “sandbox” for testing rule-based linters, prompt-based LLMs, or neural classifiers (Tonglet et al., 29 Aug 2025). These use cases reflect the dataset’s paired design: image, table, and metadata are jointly available rather than reconstructed post hoc.

Its limitations are equally explicit. The dataset covers only a restricted set of chart varieties, excluding scatterplots, maps, treemaps, and 3D pie charts (Tonglet et al., 29 Aug 2025). Each chart contains exactly one misleader, whereas real-world examples may combine two or three. Stylization is restricted to Matplotlib defaults, which induces domain bias. The dataset also does not cover “reasoning” misleaders such as deceptive titles or omitted data, nor domain-specific conventions (Tonglet et al., 29 Aug 2025).

These constraints delimit the kinds of conclusions that can be drawn from results on Misviz-synth. High performance on the benchmark would not, by itself, establish robustness to heterogeneous visualization ecosystems or to composite and semantically subtle forms of deception. Conversely, the benchmark’s controlled construction makes it valuable for isolating the visual footprints of misleading chart design, particularly when perfect supervision is required.

7. Position within the Misviz framework

Misviz-synth is positioned as the synthetic counterpart to Misviz, a benchmark of 2,604 real-world visualizations annotated with 12 types of misleaders (Tonglet et al., 29 Aug 2025). The broader motivation is that misleading visualizations are a potent driver of misinformation on social media and the web, and prior work has shown that both humans and multimodal LLMs are frequently deceived by them (Tonglet et al., 29 Aug 2025). Within that framing, Misviz-synth supplies the scale needed for training and systematic evaluation.

The relationship between the two resources is complementary. Misviz provides real-world variability and the full 12-type annotation regime, while Misviz-synth provides scale, controllability, and perfect pairing between visuals, tables, and metadata (Tonglet et al., 29 Aug 2025). In spite of its limits, the paper characterizes Misviz-synth as a uniquely large, richly annotated resource for bootstrapping models that must learn from such paired representations. A plausible implication is that the intended research workflow combines synthetic pretraining or controlled analysis on Misviz-synth with downstream evaluation on the real-world Misviz benchmark.