🎲 VSM Benchmarks — Cards & ChessImages

Two image benchmarks with very large, structured semantic spaces and fast, training-free validators, released with the paper "Score-Control for Hallucination Reduction in Diffusion Models" (VSM).

📄 Paper: https://arxiv.org/abs/2606.00377  ·  💻 Code: https://github.com/bhosalems/VSM

Both datasets are designed for systematic studies of hallucination in generative (especially diffusion) models. Each sample has a well-defined notion of validity, so a generated image can be checked — without any trained critic — for whether it lies on the data manifold (a legal board / a real card) or is a hallucination (an illegal configuration). This makes the hallucination rate a directly measurable quantity rather than a perceptual judgement.

Dataset	Domain	#Images	Resolution	Semantic space	Validity check
Cards	2×2 grids of playing cards	91,390	128×128	~10⁵	per-quadrant template match
ChessImages	Chessboards rendered from FEN	196,062	256×256	~10⁴⁴ legal states	FEN parse + python-chess legality

📦 Repository layout

Image folders are shipped as .tar archives (a handful of large files uploads far faster than hundreds of thousands of tiny PNGs); labels and metadata are plain files.

mbhosale/VSM/
├── ChessImages/
│   ├── train_images.tar        # 176,938 × 256×256 PNG  -> train_images/
│   ├── test_images.tar         #  19,124 × 256×256 PNG  -> test_images/
│   ├── train_fen.json          # {"<image_id>": "<FEN>"}  (176,938 entries)
│   └── test_fen.json           # {"<image_id>": "<FEN>"}  ( 19,124 entries)
├── Cards/
│   ├── card_imgs.tar           # 91,390 × 128×128 PNG  -> card_imgs/
│   └── templates.tar           # 40 card-face templates (64×64) for the validator
└── validators/                 # training-free hallucination checkers (see below)
    ├── chess_validator.py
    ├── cards_validator.py
    └── templates/chess/        # 26 chess-piece templates used by chess_validator.py

🃏 Cards

A Cards sample is a 2×2 grid of four standard playing cards at 128×128 (each quadrant is 64×64). The card faces are drawn from 40 base cards — ranks A through 10 across the four suits (♣ ♦ ♥ ♠); no face cards (J/Q/K). With four independently drawn quadrants the semantic space is on the order of 10⁵.

• card_imgs.tar — the full set of 91,390 grids.
• templates.tar — the 40 single-card reference templates (resized to 64×64), used by the training-free validator. Extracts to resized_templates_png/.

Hallucination check. Each of the four quadrants is template-matched (cv2.TM_CCOEFF_NORMED) against the 40 templates. If the worst quadrant match falls below a threshold, the image cannot be explained by four valid cards and is counted as a hallucination (wrong/garbled symbols, color/count errors, noise).

♟️ ChessImages

A ChessImages sample is a 256×256 rendering of a chessboard generated from a FEN string. The space of legal piece placements is astronomically large (~10⁴⁴), making this a stringent test of whether a generator stays on-manifold.

• train_images.tar (176,938) and test_images.tar (19,124) — rendered boards; filenames are <image_id>.png.
• train_fen.json / test_fen.json — map each <image_id> to its full FEN.

Hallucination check. The board image is parsed back to a FEN piece-placement string by template-matching each of the 64 squares, then its legality is tested with python-chess (chess.Board.is_valid()). Boards that violate the rules (two white kings, >8 pawns, pawns on the back rank, …) are counted as hallucinations. The validator can additionally report exact / fuzzy FEN-reconstruction accuracy against the ground-truth FENs (conditional setting).

🚀 Usage

Option A — via the VSM repo (recommended)

The VSM repo ships loaders and a downloader that fetch and unpack these archives for you:

python -m datasets.download --dataset chess --out ./data/ChessImages
python -m datasets.download --dataset cards --out ./data/Cards

Option B — directly with huggingface_hub

from huggingface_hub import snapshot_download
import tarfile, glob, os

root = snapshot_download(repo_id="mbhosale/VSM", repo_type="dataset",
                         allow_patterns=["ChessImages/*"])
for tar in glob.glob(os.path.join(root, "ChessImages", "**", "*.tar"), recursive=True):
    with tarfile.open(tar) as t:
        t.extractall(os.path.dirname(tar))
# -> ChessImages/train_images/*.png, test_images/*.png, train_fen.json, test_fen.json

✅ Verifiers (training-free hallucination validators)

The two validators used in the paper are bundled here under validators/, so you can score generated images without cloning the full code repo. They need opencv-python, numpy, tqdm, and — for chess — python-chess.

ChessImages — parse each rendered board back to a FEN and test legality with python-chess; illegal boards count as hallucinations. The piece templates ship alongside the script, so no extra flags are required:

python validators/chess_validator.py --gen-dir /path/to/generated/boards
# optional: also report FEN-reconstruction accuracy against the ground truth
python validators/chess_validator.py --gen-dir /path/gen \
    --gt-json ChessImages/test_fen.json --conditional

Cards — template-match each of the 4 quadrants against the card templates; if the worst-matching quadrant falls below the threshold, the image is a hallucination:

tar -xf Cards/templates.tar                      # -> resized_templates_png/
python validators/cards_validator.py --gen-dir /path/to/generated/cards \
    --template-dir resized_templates_png

Each validator reports the hallucination rate (H%) per folder (repeat --gen-dir for multiple seeds). Full fidelity metrics (FID / CLIP-FID / FLD) live in the VSM repo under evaluation/.

⚖️ Intended use & license

These benchmarks are released for research on generative-model reliability and hallucination. They are synthetic renderings (rule-based card grids and chess positions) and contain no personal data. Released under CC BY 4.0 — please verify this matches your intended terms before redistribution, and cite the paper below.

📑 Citation

@article{bhosale2026vsm,
  title   = {Score-Control for Hallucination Reduction in Diffusion Models},
  author  = {Bhosale, Mahesh and Devulapally, Naresh Kumar and Wasi, Abdul and
             Pham, Chau and Lokhande, Vishnu Suresh and Doermann, David},
  journal = {arXiv preprint arXiv:2606.00377},
  year    = {2026}
}

The ChessImages dataset uses python-chess and FEN strings sampled from VALUED.