Chart Deception Severity: counterfactually-augmented charts with exact ground truth, and the Deception Atlas

Companion dataset to the preprint Quantifying Chart Deception: Continuous Deception-Severity Estimation for Chart-Reading Models (Carson Rodrigues, 2026). Every chart is rendered from a real Our World in Data series twice: once honestly, once with a distortion whose parameters are set by the generator, so every label is computed from the construction rather than annotated. Contents: ground_truth_rows.jsonl (2,430 rows; 33-field machine-checkable ground-truth schema, a row populates the fields its question type defines; 636 rows carry an exact Tufte Lie Factor and severity band); train.parquet (2,129-row failure-weighted instruction-tuning projection: supported 54%, severity 27%, contradicted 13%, unanswerable 5%); holdout.parquet (114 rows); images.zip (954 chart renders, each regenerable from its render record); and the Deception Atlas (20 real published misleading charts verified against primary sources, with 44 evaluation rows and documented readings). The exact ground truth makes the severity target usable as a verifiable reward for reinforcement learning; the papers analysis plan pre-commits that use. The dataset was produced during a public model-adaptation challenge; neither the paper nor this dataset names the platform, and the deposit is scrubbed to match.