Datasets:

Abnik
/

spotdiff-v1-dev

pretty_name: SpotDiff v1 Development
tags:
  - visual-reasoning
  - spot-the-difference
  - image-to-text
size_categories:
  - n<1K
task_categories:
  - image-to-text
license: other
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/viewer/spotdiff_v1_dev.jsonl

SpotDiff v1 Development Dataset

SpotDiff is a visual spot-the-difference benchmark. Each item is one composite image containing two nearly identical panels. A model receives the complete image and identifies every visible difference using structured JSON.

This is the public development release. It contains 10 reviewed images and 61 gold differences. The annotations are intentionally public so that anyone can reproduce the evaluator locally and inspect the benchmark design.

Layout

data/images/: composite images with both panels
data/annotations/: approved golden structured differences
data/manifest.json: item IDs, image paths, and annotation paths
data/dataset_version.json: release metadata
data/viewer/spotdiff_v1_dev.jsonl: one uniform viewer row per difference

Evaluation

Install the evaluator from GitHub, or run it from a checkout:

pip install -e .
spotdiff evaluate \
  --manifest data/manifest.json \
  --predictions predictions/your-model.json

The official development score is deterministic:

0.60 * recall + 0.25 * precision + 0.15 * attribute_accuracy

Because this is a development benchmark, the answers are public and results are suitable for reproducibility and community-reported comparisons. They are not a hidden-test leaderboard result.

Image rights

Before publishing this dataset, the dataset maintainer must confirm that the source images may be redistributed. The other license marker is deliberate until the image provenance and license are documented.