vlm-modality-research/chartqa-evidence-conflict-table-v1

ChartQA-Conflict chart/table representation ablation

This derivative contains 229 audited ChartQA-Conflict items. Each row provides two visual renderings of the same official ChartQA facts:

• chart_image: the original ChartQA chart;
• table_image: a plain table image rendered from the corresponding official ChartQA CSV.

The shared question, chart-supported answer, evidence-bearing conflicting report, report-supported answer, and Source A/B assignment are identical across the two representations. The table values are not edited toward the conflicting report. This permits a controlled comparison of genuine-chart and picture-of-table evidence while holding the underlying facts fixed.

Before publication, scripts/audit_chartqa_table_ablation.py records the provenance scope and checks every row:

1. the recorded cells equal the frozen curation workbook or, when supplied, the referenced official ChartQA CSV;
2. the saved table-image checksum matches the manifest;
3. re-rendering those cells reproduces the saved table-image pixels.

One ChartQA-Conflict item (conflict_id=45) is excluded because the final report audit found that its report did not entail its designated answer. The resulting dataset contains 229 rows.

Important columns include question, chart_answer, report_answer, text_report, official_table_data, source_table, chart_image, table_image, table_image_sha256, and the frozen original-manifest hash.

This dataset inherits the use conditions and licensing requirements of ChartQA. Users should cite the original ChartQA paper and dataset.