FINGER: Fine-Grained Personalization Benchmark

FINGER evaluates recognition of visually confusable personalized concepts. This release accompanies ReCoG: Relational Concept Graph for Training-Free Personalization.

Each concept has one support image in train and separate query images in test. The original concept IDs, filenames, image bytes, support/query assignments, and split-list ordering are preserved.

Subset	Concepts	Support (train)	Query (test)	Original folders
cub_easy	20	20	380	CUB_easy/
cub_hard	20	20	380	CUB_hard/
clevr_v2	20	20	180	CLEVR/

Total: 1,000 benchmark images. There is no validation split. train is a support set for personalization, not a requirement to train the VLM.

Download the original folders

These ZIP files preserve the folder structure used by the original benchmark scripts:

Dataset	Download
CUB-Easy	CUB_easy.zip
CUB-Hard	CUB_hard.zip
CLEVR (V2)	CLEVR.zip

Each archive contains train/<concept_id>/, test/<concept_id>/, mapping.json, train_test_val_seed_23.json, and a per-file MANIFEST.csv. Additional source metadata is retained where available. Paths in the split JSON are relative to the extracted dataset folder.

For example, download all three archives with the CLI:

python -m pip install -U huggingface_hub
hf download ssun0626/FINGER --repo-type dataset --revision v1.0.0 \
  --include 'downloads/*.zip' --local-dir finger-release
mkdir -p data
unzip finger-release/downloads/CUB_easy.zip -d data
unzip finger-release/downloads/CUB_hard.zip -d data
unzip finger-release/downloads/CLEVR.zip -d data

Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("ssun0626/FINGER", "cub_easy", revision="v1.0.0")
support = dataset["train"]  # 20 images
queries = dataset["test"]   # 380 images
print(queries[0]["concept_id"])
queries[0]["image"]         # PIL image

Choose cub_hard or clevr_v2 for the other subsets. Each row has an image, an anonymous concept_id, and its original relative image_path. The Parquet files embed the original image bytes and do not require a custom loading script. Species/scene mappings are kept in separate metadata files rather than in the model input.

Protocol and metadata

• CUB-Easy contains 20 sampled CUB species; each has 1 support and 19 query images.
• CUB-Hard contains 20 visually similar CUB species; each has 1 support and 19 query images.
• CLEVR contains 20 scene concepts; each has 1 support and 9 query images.
• mapping.json maps original species names or scene IDs to anonymous concept IDs. IDs are local to each subset: the same ID may mean different species in CUB-Easy and CUB-Hard.
• Query/support images remain disjoint according to the supplied split files.
• CUB_hard/generated_info.json contains existing generated descriptions, not new manual annotations.
• Metadata copies are available under metadata/CUB_easy/, metadata/CUB_hard/, and metadata/CLEVR/. Their image paths refer to the extracted original dataset folders.
• MANIFEST.csv at the repository root describes all original dataset files inside the archives.
• SHA256SUMS covers the distributed archives, Parquet files, and metadata files.

CLEVR source version

The available local source for this release is named clevrV2, so its configuration is explicitly named clevr_v2. Its generation metadata retains the source's internal clevr_hard_v1 version string. This release does not bundle other CLEVR difficulty, hop, or rendering variants. Matching this source against the final paper's CLEVR assets has not been independently established by the packaging checks. Two auxiliary visualization PNGs are included as metadata and are excluded from the 200 benchmark images.

Source attribution and terms

CUB-Easy and CUB-Hard use images from Caltech-UCSD Birds-200-2011. The source dataset restricts image use to non-commercial research and education; image copyrights remain with their original holders. The ReCoG code license does not replace these image terms.

CLEVR images are the supplied scene renderings for this benchmark. This packaging does not assign a new license to the source assets.

For CUB-200-2011, please cite the original dataset technical report:

@techreport{WahCUB_200_2011,
  title={The Caltech-UCSD Birds-200-2011 Dataset},
  author={Wah, Catherine and Branson, Steve and Welinder, Peter and Perona, Pietro and Belongie, Serge},
  year={2011},
  institution={California Institute of Technology},
  number={CNS-TR-2011-001}
}

For ReCoG citation and evaluation instructions, see the linked project repository.