vinod-anbalagan/adaption-prism

Adaption Prism — adapted multilingual chart QA

The training set behind Prism, a LoRA for chart question answering across ten visual styles.

5,951 rows. Produced by running gridline-chartqa (3,803 hand-verified rows) through Adaption's Adaptive Data step with every recipe enabled — prompt rephrasing, metadata injection, deduplication, reasoning traces, hallucination mitigation, House Special — plus translation into four additional languages.

This is platform output, not hand-built gold data. Read the column notes below before using it. The clean source dataset is gridline-chartqa.

Languages

language	rows	share
English	3,639	61.1%
Arabic	595	10.0%
Chinese (Simplified)	593	10.0%
Spanish	571	9.6%
French	553	9.3%

Answers were translated as well as questions. 1,143 rows have non-ASCII answers — 服务, Éducation, Augmenté, 2700 دولار.

This matters: the chart images are rendered in English. A chart whose bar is labelled "Services" may carry the answer 服务. The answer is a correct translation of the correct category, but it no longer matches the text visible in the image, so exact string matching against chart content will fail on those rows. Anyone scoring against chart text should either filter to English or normalise category names back to the image's language.

Columns

column	source	trustworthy as a label?
question	original, or platform translation	yes
answer	original, or platform translation	yes, with the caveat above
enhanced_prompt	platform rewrite	rewritten, not ground truth
enhanced_completion	platform rewrite	no — see below
reasoning_trace	platform-generated	explanatory only
chart_type	source	yes
task_type	source	yes
difficulty	source	yes
notes	source	yes — includes visual style and mechanic
original_image	source	URL, see below
source, split, verified, response_constraints	mixed	metadata

On enhanced_completion. 93% of these diverge from the corresponding answer. Mean length goes from 1.4 words to 31.8 — short exact answers become paragraphs. A small number contain the platform's own preference-judging scaffolding leaked into the field, e.g. text beginning **Step 1 — Compare** | Dimension | Response 0 | Response 1. Treat this column as a record of what the platform produced, not as labels.

On original_image. These are URLs pointing into gridline-chartqa. This dataset ships no images of its own and depends on that repo staying public.

Underlying data

The source rows are 3,705 synthetic + 98 hand-authored.

Synthetic charts are rendered from seeded values, with answers computed before the image is drawn — labels never depend on reading the image. They span ten visual styles (newspaper graphics, spreadsheet defaults, journal figures, dark dashboards, dense analyst charts and others) varying colour, typography, figure geometry, title placement and grid treatment.

806 rows (21%) are perceptually hard by construction, across eight mechanics: truncated axes where bar heights misrepresent ratios, unlabeled charts with values snapped exactly onto gridlines, near-ties within 1–2%, near-identical series colours, 12–16 rotated categories, legends occluding the plot, log scales, and dual axes.

The 98 hand-authored rows come from real public dashboards — Statistics Canada, US Bureau of Labor Statistics, WHO, ECB, Bank of Canada, Climate Policy Database — each reviewed against its source figure.

Intended use

Training and evaluating short-answer chart QA, including cross-lingual settings where the question language differs from the chart language.

For clean labels, prefer gridline-chartqa. Use this one when you specifically want the multilingual rows or want to study what Adaptive Data does to a dataset.

Known limitations

• Answer/image language mismatch on translated rows, as above.
• enhanced_completion is not ground truth — 93% divergence.
• No images of its own — depends on the source repo.
• Synthetic rows come from one generator. Ten rendering styles, but a bounded set of domains, value distributions and question templates.

Provenance

Generation pipeline, style definitions and adversarial renderers: github.com/VinodAnbalagan/adaption-charts

Every synthetic row has a provenance JSON recording its seed and underlying values, so any answer can be recomputed from source.

Citation

@misc{anbalagan2026adaptionprism,
  title        = {Adaption Prism: Adapted Multilingual Chart QA},
  author       = {Anbalagan, Vinod},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/vinod-anbalagan/adaption-prism}}
}

Please also credit the upstream chart sources listed above.

---

Platform training report

Generated by Adaption AutoScientist at export time — training configuration, win-rate chart and loss curves.

This dataset is a remastered version of this dataset prepared using Adaption's Adaptive Data platform.

adaption-Prism

This dataset contains verified chart question-answer pairs focused on business, finance, and policy domains. It combines programmatically generated synthetic data with a hand-authored hardset derived from real public dashboards. The content covers various chart types and task types, requiring reasoning to extract specific values, comparisons, and trends.

Dataset size

There are 5,951 data points in this dataset. This is an instruction tuning dataset.

Quality of Remastered Dataset

The final quality is B, with a relative quality improvement of 8.7%.

Domain

• Data-analysis-visualization (100%)

Language

• English (98%)
• Chinese (2%)

Tone

• Informative (48%)
• Analytical (32%)
• Objective (16%)

Evaluation Results

• Quality Gains:

  

• Grade Improvement:

  

• Percentile Chart: