arnodjiang/MStructBench

MStructQA

MStructQA: A Multilingual Benchmark for Chart and Visual Tabular Question Answering in MLLMs

Code, prompts and evaluation · Server setup

arnodjiang/MStructBench is the dataset repository for the MStructQA project. It contains the current 24-language release: 128 base questions, 3,072 localized visuals and 8,960 distinct QA configurations. The single validation split is an evaluation split; no training split is provided. Images are localized reconstructions of charts and rendered tables, rather than upstream originals.

Load the dataset

from datasets import load_dataset

ds = load_dataset('arnodjiang/MStructBench', split='validation')
example = ds[0]
image = example['image']      # PIL image; tables are also images
question = example['query']
reference = example['answer'] # scoring only; never include in model input
context = example['source_context']  # empty when no external prose is needed

Use a pinned Hub commit via revision=... for reproducible experiments. The evaluation archive under artifacts/ contains the exact current JSONL, images, rendering code and metadata expected by the project's evaluation scripts. Download it with python -m scripts.distribution.download after cloning the code. release.json records file checksums and the canonical reference snapshot hash.

Languages and settings

English (EN), Simplified Chinese (ZH), Japanese (JA), Korean (KO), French (FR), German (DE), Spanish (ES), Portuguese (PT), Russian (RU), Arabic (AR), Hindi (HI), Italian (IT), Dutch (NL), Polish (PL), Turkish (TR), Vietnamese (VI), Indonesian (ID), Thai (TH), Swahili (SW), Persian (FA), Urdu (UR), Bengali (BN), Tamil (TA) and Telugu (TE).

• LQA: visual, question and answer use the same language: 24 configurations per base QA.
• XQA-ZH: Chinese question/answer with each of the other 23 visual languages.
• XQA-EN: English question/answer with each of the other 23 visual languages.

These are 70 distinct configurations per base QA. Aligned Chinese/English configurations are counted once in LQA. Report each LQA language separately and macro-average each XQA pivot over its 23 visual languages. AVG is the macro average over all 70 configurations, not the unweighted mean of the 26 displayed LQA/XQA columns. All language variants of a source question share a base ID.

Fields

Field	Meaning
id, base_id, case_id	Stable configuration, base QA and visual-case identifiers
image	Embedded PNG, automatically decoded by Datasets
query, answer	Current question and reference answer
query_language, image_language, answer_language	ISO 639-1 language codes
configuration	Original setting identifier; language fields define LQA/XQA membership
source_context	Translated external document prose; empty if absent
source	Upstream dataset
visual_kind	Visual category describing the primary chart type or table structure, such as Grouped Bar Chart or Column-Spanning Table
visual_family	Legacy coarse chart/table grouping for compatible evaluation cohorts
image_sha256	Checksum of the exact image bytes
metadata_json	Lossless canonical reference row, including provenance and automated audit flags

Only the question, image and optional source context are inference inputs. Do not send answers, provenance or audit annotations to the model. External prose is translated into the question language. Table contents remain in the image and are not transcribed into the prompt. For exact prompt serialization, use scripts.evaluation.context_input.input_text on json.loads(metadata_json).

Visual categories

visual_kind is assigned through automated visual classification of one current English image per base case and shared across its localized variants. It describes the visible chart type or table merge structure; questions and answers are not classifier inputs. Tables use Simple Table, Row-Spanning Table, Column-Spanning Table or Mixed-Spanning Table. Simple means no merged rows/columns, not a one-cell table. Charts use a more specific vocabulary for bars, lines, distributions, spatial fields, diagrams and composites. visual_family retains the legacy chart/table cohort. See taxonomy definitions. The exact archive includes visual_taxonomy.json and visual_classification.json with per-case evidence, secondary types, layout and model confidence. These are model-generated annotations, not human certification. The change affects metadata only: images, queries, reference answers and source context are unchanged.

Sources and construction

Source	Base questions	QA configurations
CharXiv	48	3,360
ChartQAPro	27	1,890
TableVQA-Bench	22	1,540
ChartQA	12	840
Visual-TableQA	10	700
MMTU	9	630

The pipeline selects source-linked QA, reconstructs visuals, localizes labels and linked QA, conservatively rewrites questions, renders images, and verifies numerical structure and source alignment. Five source cases include external prose (four MMTU/FinQA and one ChartQAPro), localized across all 24 languages; these affect 350 QA configurations. Current-only exports omit historical questions, previous labels, old data versions and model predictions. Source revision, row and file identifiers are retained for attribution and tracing.

Evaluation

First apply deterministic answer matching. Non-matching predictions are judged for answer equivalence by a separately configured, text-only LLM judge. Only equivalent receives credit; different and uncertain outcomes are incorrect. Failed/missing predictions count as incorrect in the fixed denominator. Record the inference and judge models, prompts, token usage, request counts and release revision.

Quality and limitations

This release contains the full candidate cohort, including records flagged needs_review by automated checks; it is not a claim that every record passed human verification. Inspect metadata_json for audit status. Reconstruction, translation, source ambiguity and script rendering may affect results. Automated review is not independent human certification. The 128 base questions and their correlated language variants are not 8,960 independent source observations. Upstream overlap and potential model exposure should be considered when interpreting scores. No original-image baseline is implied by these results.

Licensing

The code repository uses MIT. Dataset examples remain subject to their respective upstream terms; MIT does not apply to the combined dataset. The other license tag denotes source-specific terms, not a new blanket grant. Consult the six linked upstream dataset cards and the retained per-example provenance before reuse or redistribution. Cite the applicable upstream sources alongside MStructQA.

Citation

@misc{mstructqa,
  title = {MStructQA: A Multilingual Benchmark for Chart and Visual Tabular Question Answering in MLLMs},
  howpublished = {\url{https://github.com/arnodjiang/MStructQA}},
  year = {2026}
}

This is a repository citation; no accepted venue or publication DOI is asserted.

Observed visual types

The current release contains 32 primary types. Counts are base cases; each case contributes 70 QA configurations.

Type	Base cases	QA
Multi-Series Line Graph	23	1610
Simple Table	20	1400
Column-Spanning Table	15	1050
Mixed Chart	9	630
Grouped Bar Chart	5	350
Line Graph with Uncertainty Bands	5	350
Heatmap	4	280
Line Graph	4	280
Pie Chart	4	280
Horizontal Bar Chart	3	210
Mixed-Spanning Table	3	210
Stacked Bar Chart	3	210
Area Chart	2	140
Bar-Line Combination Chart	2	140
Chart-Table Composite	2	140
Cumulative Distribution Plot	2	140
Density Plot	2	140
Diverging Bar Chart	2	140
Line Graph with Error Bars	2	140
Phase Diagram	2	140
Row-Spanning Table	2	140
Scatter Plot with Error Bars	2	140
3D Streamline Plot	1	70
Bubble Chart	1	70
Confusion Matrix	1	70
Contour Plot	1	70
Correlation Matrix	1	70
Histogram	1	70
Lollipop Chart	1	70
Scatter Plot	1	70
Table-Diagram Composite	1	70
Vertical Bar Chart	1	70