sungguk/visual-answerability

Visual Answerability across Charts and Scenes

Frozen release: answerability-v1.0.0-20260928, Hub tag v1.0.0. Dataset: sungguk/visual-answerability. This benchmark accompanies When Does an Image Determine the Answer? Benchmarking Visual Answerability across Charts and Scenes, using the September 28, 2026 manuscript metadata and the unchanged evaluated cohorts. No arXiv identifier or DOI has been assigned in this package.

Each source question has several original/edited views. Present one image and one question independently to a model; group responses only after inference. The main task requires correct supported answers and correct abstentions for every view of a question. Do not expose state, target, answerable, sibling views, or evidence files to the model.

Contents and access

Configuration	Groups / views	Image access
plotqa	1,000 / 5,000	Embedded PNG
clevr	1,000 / 4,000	Embedded PNG
gqa	1,000 / 3,000	Local reconstruction

All three configurations use an evaluation-only test split. Source splits remain explicit: PlotQA comes from the frozen train-prefix selection; CLEVR and GQA come from source validation data. This naming does not imply use of an official upstream hidden test split. This release covers the 12,000 evaluated views, not the broader 58,000-view training/development resource.

The chart cohort uses executable D15 direct-lookup questions from the frozen PlotQA prefix. CLEVR uses source functional programs; GQA uses the supported native semantic-program subset. There is one source image per question group within each domain.

The gqa configuration supplies IDs, our labels and edit recipes. Its image and question fields are null. The official GQA pages and archive READMEs checked for this preparation did not supply an explicit annotation redistribution grant. The earlier local CC BY declaration is not treated as upstream permission. Obtain the official questions, graphs and images yourself, then follow RECONSTRUCTION.md. The reconstruction checks source identity, programs, labels and exact evaluated image bytes. It writes outside this folder.

PlotQA and CLEVR data retain CC BY 4.0; bundled project code retains Apache 2.0. See LICENSE.md and ATTRIBUTION.md for component boundaries, upstream authors, modifications and notices.

Load the dataset

python -m pip install datasets pillow

from datasets import load_dataset

root = "sungguk/visual-answerability"
charts = load_dataset(root, "plotqa", split="test", revision="v1.0.0")
scenes = load_dataset(root, "clevr", split="test", revision="v1.0.0")
gqa_index = load_dataset(root, "gqa", split="test", revision="v1.0.0")
example = charts[0]
example["image"].save("example.png")
print(example["question"])

For the evidence, scoring scripts and pinned dependencies, download the full snapshot as described in UPLOAD.md. The v1.0.0 tag identifies the verified release commit; use that commit hash when an immutable revision is required. Local snapshot paths can also be passed to load_dataset.

Fields and states

item_id is the stable prediction join key (item-0001… or cross-00000…). group_id joins sibling states. source_question_id, source_image_id, source_split, and source_revision preserve upstream identity. question_sha256 and image_sha256 validate reconstruction and release integrity; hashes are not the semantic proof of a label. image is a Hugging Face Image feature backed by embedded bytes, except for GQA. target is a string, with <UNANSWERABLE> for unanswerable states. dependency_depth is the historical field name; for scene programs it counts steps, not tree depth.

• FULL: the original supported question.
• A_SAME: a control edit preserves the supported answer.
• A_CHANGED: an edit changes the supported answer (charts and CLEVR).
• U_MISSING: the construction removes evidence needed for a unique answer.
• U_INVALID: an edit invalidates the chart question's referent (charts only).

evidence_kind, evidence_file, operator, source_view_id, and gqa_location_stratum describe the construction and its evidence. The location flag identifies 376 GQA groups for the paper's sensitivity analysis; it is not a finding that those labels are wrong or that the remaining 624 groups are clean.

Evidence and evaluation

• Chart proofs: evidence/plotqa/bounded-witness-proofs.jsonl.gz contains all 1,000 corrected bounded witnesses, reconstructed from frozen FULL and A_CHANGED worlds and verified against each evaluated missing-view PNG. The checker re-renders complete worlds, verifies axis admissibility and two executors, then checks identical intervened pixels with different answers. These are freshly serialized proofs, not a claim of byte identity with an unavailable historical proof archive. Original per-view certificate descriptors are retained as historical construction metadata. The corrected chart missing-information claim is verified from the separate bounded-witness records above.
• Scene evidence: CLEVR programs, worlds, alternatives and render/edit metadata are in evidence/clevr/. GQA source evidence is reconstructed from official inputs using evidence/gqa/reconstruction.jsonl.gz. Neither scene subset is presented as having the chart complete-world pixel guarantee.
• Scoring records: inputs/ contains six configurations' 72,000 selected final scoring projections, including 11 terminal invalid outputs. These are parsed scoring fields, not raw generations, token traces or fresh inference. GQA question text is replaced by stable question-ID references; its rationale text is omitted. Scoring and all reported numerical results are preserved. Image paths inside the historical scoring manifests and group descriptors describe the original construction layout; use the Parquet image column and stable item_id for public model inputs.
• Statistics: exact replay of all 18 model/domain cells and 10,000-draw paired group bootstrap intervals; also saved finite-program records for 1,024 original and 1,456 combined solver-study cases. Replaying these records is separate from rerunning the solver constructors.

See EVALUATION.md for new predictions and replay commands, PROTOCOL.md for prompts and serving configurations, and UPLOAD.md for snapshot download and release verification.

Intended use and limitations

Use for evaluation of image-conditioned answers and abstention, construction audits, and reproduction of reported analyses. Do not train or tune on this evaluation split when reporting its benchmark results. Questions were selected by deterministic executable-program and rendering gates, not representative sampling of all visual questions. New per-example edits/labels used no human adjudication or reasoning-model judgments.

The chart claim assumes correct source truth, question/program compilation, the fixed-axis rational-valued world class, and the drawing primitives. The two executors are independent implementations; they share rendering code. CLEVR labels use source programs and edits. GQA graphs and object boxes may be incomplete; position questions can retain visible cues. Report GQA scores as agreement with these source-derived labels, including both sensitivity strata. There is no independent human adjudication of every view. Configuration comparisons include different serving backends; Qwen 27B scenes mix preserved API outputs and a local FP8 fallback. Intervals are descriptive, paired by question, and not corrected for multiple comparisons.

Upstream photographs may depict people and everyday places. The release adds no identity annotation or sensitive-attribute classification. All original rights and any third-party rights remain with their respective holders.

Citation and provenance

Use CITATION.cff or CITATION.bib, and cite the upstream datasets in ATTRIBUTION.md. The author list follows the prepared manuscript; no conference acceptance or published identifier is asserted. RELEASE.json records the build scope and source snapshot. MANIFEST.json identifies every frozen file, including the actual working-tree code used despite existing uncommitted manuscript revisions. Reports under validation/ state what was actually checked during local preparation. Their historical publication_performed: false fields describe those offline checks, not the current Hub visibility. The release retains the exact evaluated data, proofs, scoring records and software of the frozen preparation candidate.