README.md

metadata

language:
  - ru
  - en
license: other
license_name: mixed-research-use
size_categories:
  - 1K<n<10K
task_categories:
  - visual-question-answering
  - question-answering
task_ids:
  - visual-question-answering
pretty_name: RuChartQA
tags:
  - chart-understanding
  - visual-reasoning
  - russian
  - benchmark
  - vlm-evaluation
configs:
  - config_name: chartbasic
    data_files: synthetic/chartbasic.jsonl
  - config_name: chartreasoning
    data_files: synthetic/chartreasoning.jsonl
  - config_name: chartperception
    data_files: synthetic/chartperception.jsonl
  - config_name: chartreal
    data_files: chartreal/data.jsonl

RuChartQA

A Russian-language chart question answering benchmark for evaluating Vision-Language Models, with both synthetic and real-world evaluation sets.

Dataset summary

Split	Examples	Charts	Source
Synthetic ChartBasic	360	90 (×4 variants)	Generated
Synthetic ChartReasoning	480	120 (×4 variants)	Generated
Synthetic ChartPerception	360	90 (×4 variants)	Generated
ChartReal	242 QA	96 charts	Rosstat, Bank of Russia (PDF)
Total	1442 QA	396 unique charts

The synthetic split has 4 variants per chart: ru_image, en_image, ru_text (text description instead of image), en_text — enabling controlled language and modality ablations. ChartReal is ru_image only.

Why this benchmark

Most chart-QA benchmarks (ChartQA, PlotQA, FigureQA) are English-only. Existing Russian-language chart evaluation has been limited to translated subsets. This benchmark addresses two gaps:

1. Language coverage. Native Russian questions, Russian-language axis labels, captions, and currencies (₽).
2. Real-world distribution shift. Synthetic-only benchmarks systematically overestimate VLM performance on real-world graphs from government statistics and central bank publications. Our analysis (see results/leaderboard.csv and the [accompanying paper]) shows performance gaps of +11 to +41 percentage points between synthetic and real-world splits across three modern VLMs.

Loading

from datasets import load_dataset

# Real-world split (the one with the bigger story)
chartreal = load_dataset("romath/RuChartQA", "chartreal")

# Synthetic splits
chartbasic = load_dataset("romath/RuChartQA", "chartbasic")
chartreasoning = load_dataset("romath/RuChartQA", "chartreasoning")
chartperception = load_dataset("romath/RuChartQA", "chartperception")

Schema

Each row has:

Field	Type	Description
example_id	string	Unique identifier (e.g. chartreal_007_q2_ru_image)
subdataset	string	ChartBasic, ChartReasoning, ChartPerception, or ChartReal
variant	string	ru_image, en_image, ru_text, en_text
language	string	ru or en
modality	string	image or text
chart_type	string	bar, line, mixed, pie
chart_id	string	Chart identifier (multiple QA may share one chart)
question_type	string	lookup, comparison, min, max, difference, conditional
question	string	Natural-language question
answer	string	Gold answer
answer_numeric	float | null	Numeric form if applicable (for tolerance scoring)
answer_type	string	numeric or categorical
image_path	string | null	Relative path to PNG (for image variants)
text_description	string | null	Text description of the chart (for text variants)

Evaluation

We provide a normalizer (eval/normalize.py) that handles:

• Numeric tolerance (5%, the ChartQA standard) with a year-as-numeric exception requiring exact match (1900–2100)
• Bidirectional substring matching for categorical answers (gold ⊆ pred or pred ⊆ gold), disabled when gold contains compound markers (и, or, ,)
• Lower/strip/punctuation normalization

Minimal example:

python3 eval/eval_example.py predictions.jsonl chartreal/data.jsonl

A prediction file is JSONL with {"example_id": ..., "prediction_raw": "..."} per line.

Baselines

Predictions on ChartReal from four systems are included in baselines/:

System	ChartReal Accuracy	Synthetic ru_image
Qwen3-VL 32B Instruct	75.2%	86.3%
Gemini 2.5 Flash	71.1%	92.7%
Nemotron Nano 12B v2 VL	45.9%	86.7%
OCR + Llama 3.3 70B (text-only baseline)	34.7%	n/a

All gaps between systems on ChartReal are statistically significant (95% bootstrap CI) except Qwen vs Gemini (Δ=+4.1pp, CI [−1.2, +9.5], p=0.16). See results/leaderboard.csv.

Construction

Synthetic

Generated from category templates (cities, products, demographics, etc.) with controlled distributions over chart types (bar) and question types. Each chart was rendered in Russian and English; for each language, both an image and a text-description variant exist. This 4-way structure allows clean ablations of language and modality effects.

ChartReal

Charts were extracted from public PDF reports of:

• Rosstat (Russian Federal State Statistics Service) — annual and monthly statistical bulletins
• Bank of Russia (CBR) — financial stability reports, monetary policy commentary

Each chart received 1–4 questions covering different reasoning types. Charts span four types (bar, line, mixed, pie) with realistic noise: small fonts, dense legends, multi-axis scales, and stylistic conventions specific to Russian government publications.

Licenses

This dataset uses mixed licensing:

• Code (eval/normalize.py, eval/eval_example.py): Apache 2.0
• Synthetic QA + images (synthetic/): CC-BY 4.0 — author's original work
• ChartReal QA annotations (chartreal/data.jsonl): CC-BY 4.0 — author's original annotations
• ChartReal images (chartreal/images/): research use only, original copyright preserved. These are derivative works (PNG renderings of pages from public-domain government PDFs). Original publishers (Rosstat, Bank of Russia) retain copyright on the visual material. Re-use beyond academic research may require permission from the original publishers.

By using the chartreal/images/ portion, you agree to:

1. Use it only for academic / non-commercial research
2. Cite both this dataset and the original publisher
3. Not redistribute the images independently of the QA annotations

Citation

@dataset{ruchartqa_2026,
  title  = {RuChartQA: A Russian-Language Chart Question Answering Benchmark with Synthetic and Real-World Splits},
  author = {Roman <last name>},
  year   = {2026},
  url    = {https://huggingface.co/datasets/romath/RuChartQA},
  note   = {HSE Bachelor's thesis}
}

Limitations

• ChartReal is image-only. A text_description variant for real-world charts is not provided — automatic transcription of complex line/mixed charts to faithful text without losing information turned out to be infeasible in practice.
• Bar-bias in synthetic. All synthetic charts are bar-type. Comparison fairness across chart types should use the bar-only subset of ChartReal (n=67) — see results/leaderboard.csv.
• Answer normalizer judgement calls. A small number of answers (≤2pp of total) are influenced by language-drift conventions: yes/no in English vs Russian, Roman vs Cyrillic month numerals. We chose conservative scoring (mismatch counted as wrong); reasonable alternatives exist.

Contact

Questions, errata, or contributions: [your email or GitHub username].