anonymous-submission-2026/SemanticStress

Dataset Card

Purpose

SemanticStress evaluates whether vision-language models choose image-faithful captions over semantically stressed alternatives that alter one controlled aspect of meaning.

Code And Reconstruction

The anonymous code release is available at https://anonymous.4open.science/r/SemanticStress-1234/.

This Hugging Face Dataset is a metadata-only release. It intentionally omits image bytes, private local paths, and original source/base captions. To build the complete local evaluator-ready dataset, clone the anonymous code release, obtain the upstream image assets under their licenses, and run the reconstruction workflow documented in the repository README.

Minimal reconstruction command:

python scripts/evaluation/prepare_hf_split_for_eval.py \
  --repo-id anonymous-submission-2026/SemanticStress \
  --split benchmark_core_v1 \
  --output-parquet data/hf_eval/benchmark_core_v1.parquet \
  --image-root /path/to/coco/val2017 \
  --image-root data/source_images/aro_hf \
  --image-root data/source_images/winoground \
  --missing-report reports/hf_eval_missing_images_core.csv

Run the same command with --split analysis_subset_v1 and --output-parquet data/hf_eval/analysis_subset_v1.parquet to reconstruct the analysis subset. The script restores local image_path values and fills the evaluator-required base_caption field from the released positive_caption/preserve_caption fields.

Source Datasets

SemanticStress is built from three public source families used in the paper: SugarCrepe, ARO, and Winoground. SugarCrepe image references point to COCO Val2017 images. Source images and original source/base captions are not redistributed in the public Hugging Face metadata splits. Release users should consult the final paper and upstream licenses for the exact released data bundle and attribution terms.

Structure

Each public example has a preserve caption and a stress caption. The preserve caption should remain faithful to the image. The stress caption should be fluent but semantically mismatched. Internal construction also tracks a source/base caption, but that field is omitted from public metadata exports.

Public Hugging Face exports do not embed image bytes. They retain anonymized relative image names in image_path, for example images/000000000000.jpg, plus source identifiers needed to reconstruct local image paths after users obtain the upstream images under their licenses. Private local paths and intermediate cache paths are not included.

Core fields are sample_id, variant_id, image_id, image_path, source, subset, stress_type, evaluation_slice, stress_level, stress_score, preserve_caption, stress_caption, positive_caption, negative_caption, perturbation_operation, target_slots, and preserved_slots.

positive_caption is the pairwise evaluation alias for the image-faithful caption. negative_caption is the pairwise evaluation alias for the stressed caption; in this release it is equivalent to stress_caption.

Stress Types

The main stress types are object, attribute, relation, and compositional. Examples are grouped into five diagnostic evaluation slices: object, attribute, relation, order-compositional, and high-confusion compositional.

Operation Metadata

Rows may include perturbation operation metadata, target and preserved slots, source identifiers, and feature metadata used by rule-based scoring.

Severity Levels

Severity is assigned from the continuous stress score:

• mild: score <= 0.42
• moderate: 0.42 < score <= 0.62
• strong: score > 0.62

Benchmark Core And Analysis Subset

The public Hugging Face dataset exposes two splits:

• benchmark_core_v1: the main evaluation set.
• analysis_subset_v1: a smaller matched set for controlled probing, audit, and repeated API runs.

Validation Summary

Validation utilities support human pilot summaries, human/LLM alignment, LLM-as-judge audits, and severity-level diagnostics. The release repo does not include private annotator identifiers.

Intended Uses

Use SemanticStress to measure semantic robustness, order sensitivity, prompt sensitivity, repeat sensitivity, invalid response rates, and slice-level weaknesses of VLMs.

Non-Intended Uses

Do not use the benchmark as a complete measure of visual understanding, as a substitute for task specific safety evaluation, or to identify private individuals.

Known Limitations

Stress captions are controlled approximations of semantic perturbations and may not cover all linguistic or visual phenomena. Some source slices may be smaller than others, and severity is a rule-based proxy rather than a human-perfect scale.

Licensing

Code is released under the license in LICENSE. Public metadata is provided without redistributing upstream image files or original source/base captions. Users are responsible for obtaining COCO, SugarCrepe, ARO, and Winoground source assets under the applicable upstream licenses before reconstructing a local evaluator-ready copy.