stellalisy/VVRBench

VVR-Bench

VVR-Bench evaluates visual instruction following with deterministic, model-free verifiers organized into five families: Grounding, Cardinality, Spatial, Size, and Topology.

The repository contains one complete frozen verifier library at vvr_bench/verifier.py. It implements all 46 atomic verifications used by the project, including repeated-group direction, alignment, one-to-one containment, and relative-size requirements.

The default Core configuration contains 10,000 prompts. Its five approximately equal-mass complexity bins contain 2,082, 1,925, 1,988, 1,998, and 2,007 prompts.

The Challenge configuration contains 720 jointly satisfiable prompts: twenty prompts at every integer complexity level from 45 through 80. It covers all 42 structural subtypes, balances structural complexity across Cardinality, Spatial, Size, and Topology, and was constructed independently of the Core prompt set. Every released Challenge specification has a programmatically rendered full-scene strict pass and a targeted counterfactual failure.

The Fast configuration contains 820 prompts sampled from Core. It contains 20 prompts at every attainable integer complexity level from 3 through 44 (level 5 is unattainable under the frozen complexity definition). Use VVR-Bench-Fast for rapid development and model comparisons; use the complete Core configuration for final benchmark reporting.

Each row contains the natural-language prompt, formal visual specification, active atomic criteria, composition metadata, and continuous and discrete difficulty metadata. It does not contain generated images or model outputs. Evaluator version 1.4.2 derives all reported family names and composition counts from the atomic criteria using the canonical taxonomy; it does not use the older construction-group labels.

Generate Images

Generate one image per row and name it {id}.png, for example vvr_bench_00000.png. Images may also be supplied through a JSONL manifest containing id and image_path fields.

Evaluate

Clone this dataset repository and install its evaluator:

git clone https://huggingface.co/datasets/stellalisy/VVRBench
cd VVRBench
pip install -e .
vvr-bench-evaluate \
  --predictions-dir /path/to/generated_images \
  --output-dir /path/to/results \
  --require-complete

Core remains the default. Evaluate the Challenge configuration with:

vvr-bench-evaluate \
  --config challenge \
  --predictions-dir /path/to/challenge_images \
  --output-dir /path/to/challenge_results \
  --require-complete

For the 820-prompt development subset, replace challenge with fast.

All configurations are also available through datasets:

from datasets import load_dataset

core = load_dataset("stellalisy/VVRBench", split="test")
challenge = load_dataset("stellalisy/VVRBench", "challenge", split="test")
fast = load_dataset("stellalisy/VVRBench", "fast", split="test")

The evaluator writes summary.json and per_example.jsonl. Report strict_accuracy, the fraction of prompts whose image satisfies every constraint, overall and by complexity range (by_difficulty). dense_vvr is a partial-credit score and must not be reported as accuracy.

The evaluator can download the benchmark Parquet automatically. Pass --data to use a local copy instead.

Citation

@misc{li2026verifiablevisualrewards,
  title  = {Verifiable Visual Rewards Transfer from Synthetic Scenes to Natural Prompts},
  author = {Shuyue Stella Li and Xiaochuang Han and Yulia Tsvetkov and Luke Zettlemoyer},
  year   = {2026}
}

Reproducibility

• Core version: 1.3
• Challenge version: 2.0
• Fast version: 1.0
• Evaluator revision: 1.4.2
• Immutable data revision: 7a79a160060e15a9af52c3737e0ad1a83186804c
• Core metadata source SHA256: 8e506931924c89718bce9c8a9d869abc38b34d6ab70e155a529bc26063f04b26
• Core Parquet SHA256: c246cabb6e20e6de5b32c7b4054eac144144cdd18869dcb29791ced09476c89f
• Challenge metadata source SHA256: 18672786123c0630791488411e0a6e94e90e6f8eb6b015f2e0d24488beb49519
• Challenge Parquet SHA256: 3b6c6baf5318842e5a113116fe4ff28a0442b70c35b38669d3dbb28d57989c45
• Fast data revision: 51c69da829f13fe705b8393606e22d84b87b8177
• Fast Parquet SHA256: 2cf0fac7671d74fd01f5ba5aca435e79755f39f323b206260c031c42a59ab035
• Frozen verifier SHA256: 2c3661be030c078e69c53284949524b4473150c67fd23bb294ffaafd7fa64f77

See VERIFIER.md for the complete contract and validation record.

Training data

The vvr_easy and vvr_matched configurations hold the RLVVR training sets, 100,000 tasks each, generated independently of the benchmark:

• vvr_easy: each task adds one constraint type from one generation stratum; structural complexity at most 20.
• vvr_matched: matched to the VVR-Bench complexity distribution and generation strata.

No training prompt or task specification appears in the Core, Fast, or Challenge sets. Each row has the prompt, its structural complexity (complexity, the same vvr_complexity_v8 score as the benchmark), and the formal visual specification, complexity breakdown, and generation metadata as JSON strings (visual_spec_json, complexity_json, training_data_json). Training code: https://github.com/stellalisy/VVRBench

from datasets import load_dataset

vvr_easy = load_dataset("stellalisy/VVRBench", "vvr_easy", split="train")