pandeyankit84/autoscientist-chartqa-dataset

AutoScientist Chart-QA Dataset

A chart-understanding dataset for the Adaption AutoScientist Challenge (Data Visualization). Its distinguishing features: answers are correct by construction (computed from the underlying chart data, not human-labeled) and it includes a Hindi/Devanagari + romanized slice with matched en/hi twins for a clean cross-language comparison.

• Published set: 471 rows (HF + Kaggle). Regenerate any size with the reproduction command below.
• HF: https://huggingface.co/datasets/pandeyankit84/autoscientist-chartqa-dataset
• Kaggle: https://www.kaggle.com/datasets/pandeyankit99/autoscientist-chartqa-dataset

Composition

Each row is one (chart image, question, answer). Splits: train / val / test, plus test_novel (chart types held out entirely, to measure generalization). Indic en/hi twins share a pair_id and are kept within the same split so no twin leaks across the train/test boundary. Per-source/type/kind/lang breakdown is in stats.json.

Chart types: bar, hbar, grouped_bar, stacked_bar, line, multiline, pie, scatter, area. QA kinds: value_lookup, max, min, compare, sum, count, trend, proportion, difference, mean. Languages: en, hi (Devanagari), hi-romanized.

How answers are guaranteed correct

Every value drawn on the chart is produced by the same formatter as the gold answer, and all ground-truth is computed with integer-quantum arithmetic (exact tie/threshold handling), unique-extremum guards, Pearson-r for scatter correlation, and pie proportions labeled from the computed percentages. See autoscientist_toolcaller/viz/synth_charts.py (invariants I1–I7) and the offline test suite tests/viz/test_viz.py.

Provenance & licensing

Synthetic and Indic charts are original, released under Apache-2.0. Optional ReachQA rows are MIT (attribution preserved). Devanagari rendering uses OFL Noto (or system) fonts. No scraped or non-commercial data — clean for the mandatory Hugging Face + Kaggle release.

Intended use

Supervised fine-tuning of vision-language models for chart reading and reasoning, including multilingual (Hindi) chart understanding. See the companion model card (VIZ_MODEL_CARD.md).

Limitations

• matplotlib-styled synthetic charts (clean); complements but doesn't replace real-world chart corpora (e.g. CharXiv-style dense arXiv figures).
• Numeric answers are graded with ±5% relaxed tolerance.

Reproduction

python -m autoscientist_toolcaller.viz.build_dataset --out data/viz --n-synth 400 --n-indic 200 --seed 42