VisionValueBench

VisionValueBench studies visual value leakage in multilingual text-to-image generation. Project homepage · Code and execution prompts

Scope

This package contains the 2,773 retained scenarios, their 960-scenario evaluation subset, and 68,764 evaluation-image records. The frozen benchmark contains 68,764 completed IDs against a design target of 69,120 across six base generators and two extensions.

Evaluation JPEG bytes are preserved exactly as submitted to the automatic judges and checked against recorded review-image hashes. Scenario text and frozen numerical labels accompany each image. Final annotation files contain binary votes and decisions; this package does not include rich evidence responses or original generated images.

Configuration	Records
flux2	5,760
flux2_upsampler	5,760
gptimage2	13,243
hidream	5,760
hidream_refiner	5,760
ideogram4_magic	5,753
mai26flash	13,302
nanobanana2	13,426

Scenario collections

Configuration	Split	Scenarios	Scope
scenarios_2773	corpus	2,773	Full retained scenario collection after removing 180 of the 2,953 constructed scenarios
evaluation_scenarios_960	test	960	Final evaluation subset, with 160 scenarios per value dimension

Both tables contain the canonical English scenario text, scenario_id, value_dimension, opposing value_endpoints, concept, construction_scenario_id, and is_evaluation. The same scenario_id joins these tables to the image records and numerical labels. Every evaluation scenario belongs to the retained collection. The final evaluation set retains 897 initially selected scenarios and adds 63 replacements after screening. The 63 replacements were sampled using refill seed 20260921.

from datasets import load_dataset
scenarios = load_dataset("ZucanLyu/VisionValueBench", "scenarios_2773", split="corpus")
evaluation_scenarios = load_dataset("ZucanLyu/VisionValueBench", "evaluation_scenarios_960", split="test")

The corpus and test names describe these collections; they do not define a disjoint training/test partition. These tables can be loaded without downloading images. JSONL downloads are also available: 2,773 scenarios and 960 evaluation scenarios. SCENARIO_SPEC.json, SCENARIO_PROVENANCE.json, and SCENARIO_ARTIFACT_MANIFEST.json document the fields, frozen source hashes, screening, and verification. The canonical English text does not include the translated generation prompts.

Fields and judgments

sample_id, model_configuration, scenario_id, prompt_language, culture_cue, and six_shared_conditions identify each scientific condition. scenario holds the English scenario context supplied to the evaluators; it is not the translated generation prompt. image contains embedded image bytes with a relative sample filename; evaluation_sha256, image_width, and image_height describe those bytes.

SAE (Scenario-Anticipated Evaluation) and IAE (Image-Adaptive Evaluation) remain separate. Each method has three judges (luna, gemini, muse) and 12 binary endpoint votes. sae_votes and iae_votes retain each judge's vector; sae_supported and iae_supported are per-endpoint majorities (at least two supporting judges). Opposing orientations are judged independently. See dataset_spec.json for endpoint order and judge model names.

Load

from datasets import load_dataset
dataset = load_dataset("ZucanLyu/VisionValueBench", "all", split="test", streaming=True)
row = next(iter(dataset))
image = row["image"]

The test split names the evaluation collection; it does not imply an independent training/test partition. Each named configuration selects its own shards; all references the same files without duplicating records.

The same numerical labels are available separately in labels.jsonl.gz for analyses that do not need image downloads.

Verification and availability

ARTIFACT_MANIFEST.json records shard hashes, sizes, row counts, and image-byte totals. BUILD_STATUS.json records local build/readback checks. Repository access and viewer availability depend on the current Hub settings. Licensing has not been specified.