evaluation_code/README.md

VWG-Bench evaluation toolkit

This toolkit validates VWG-Bench metadata and evaluates generated videos with the benchmark's VLM-as-Judge protocol.

Installation

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .

Set judge credentials through the environment:

export GEMINI_API_KEY="..."
export VWG_JUDGE_MODEL="gemini-2.5-pro"

Credentials are never loaded from source-controlled files.

Expected generated-video layout

By default, evaluation looks for:

videos/
├── 0_seed0.mp4
├── 0_seed1.mp4
├── 0_seed2.mp4
└── ...

Use --filename-template for another convention. Available placeholders are {id}, {id06}, and {seed}.

Validate the dataset

vwg-bench validate-data \
  --dataset-root /path/to/VWG-Bench

This verifies metadata fields, IDs, all 380 images, dimensions, 38 task groups, image shapes, and SHA-256 hashes.

Evaluate VWG-Bench

bash scripts/eval_vwg.sh \
  /path/to/VWG-Bench \
  /path/to/videos \
  outputs/model_name/results.jsonl \
  0,1,2

The evaluator samples at most 16 frames, includes the true final frame, and computes applicable 1–5 metrics:

• video quality;
• progress consistency;
• implicit-rule following;
• progress-goal realization;
• last-frame-goal realization.

Metrics without an applicable annotation are omitted rather than assigned zero. Results are resumable by result_id.

Other reported benchmarks

Four cleaned entry scripts are provided:

scripts/eval_vwg.sh
scripts/eval_mme_cof.sh
scripts/eval_ruler_bench.sh
scripts/eval_v_reasonbench.sh

MME-CoF uses the local five-aspect VLM evaluator. RULER-Bench and V-ReasonBench are format adapters only and require pinned upstream repositories. See external_benchmarks/README.md before reporting results.

Release verification

bash scripts/validate_release.sh /path/to/VWG-Bench

License

The evaluation toolkit is released under CC BY-NC 4.0. See LICENSE.md.