VBVR-Test-Bench

A reorganization of the Video-Reason/VBVR-Bench-Data test set, organized first by split (In-Domain / Out-of-Domain), then by class (Instruction_Following / Pure_Reasoning), and inside Pure_Reasoning, further split into unchanged vs rewritten.

Class / bucket	Tasks	Samples
Instruction_Following	56	280
Pure_Reasoning / unchanged	27	135
Pure_Reasoning / rewritten	17	85
Total	100	500

rewritten tasks are reasoning tasks whose upstream VBVR prompt contained answer leaks, algorithm-name hints, redundant procedural decomposition, or pure rendering choreography. We rewrote these prompts to test the model's reasoning rather than its ability to mimic the spelled-out solution. Both the rewritten prompt.txt and the original VBVR prompt_original.txt are included so you can audit each diff.

Layout

VBVR-Test-Bench/
├── README.md
├── In-Domain_50/
│   ├── Instruction_Following/
│   │   └── <task>/<sample>/
│   │       ├── prompt.txt
│   │       ├── first_frame.png
│   │       ├── final_frame.png
│   │       ├── ground_truth.mp4
│   │       └── metadata.json
│   └── Pure_Reasoning/
│       ├── unchanged/
│       │   └── <task>/<sample>/...    # prompt.txt = upstream VBVR
│       └── rewritten/
│           └── <task>/<sample>/
│               ├── prompt.txt           # rewritten (leak-stripped)
│               ├── prompt_original.txt  # upstream VBVR (reference)
│               ├── first_frame.png
│               ├── final_frame.png
│               ├── ground_truth.mp4
│               └── metadata.json
└── Out-of-Domain_50/
    ├── Instruction_Following/...
    └── Pure_Reasoning/
        ├── unchanged/...
        └── rewritten/...

Source

• Prompts (prompt_original.txt and Instruction_Following / Pure_Reasoning/unchanged prompt.txt): byte-identical to upstream Video-Reason/VBVR-Bench-Data.
• Videos / frames / metadata: byte-identical to upstream VBVR-Bench-Data.
• Pure_Reasoning/rewritten/<task>/<sample>/prompt.txt: derived from prompt_original.txt by applying our rewrite rules.

What got rewritten

The 17 rewritten tasks (see prompt.txt vs prompt_original.txt for the exact diff per sample):

Choreography removal (5 tasks)

• G-13, G-16, G-41, G-45, G-47: stripped trailing showing the complete movement process step by step (pure rendering instruction).

Answer / leak removal (8 tasks)

• G-131, G-134, G-135: the pattern string spelled the answer literally (e.g., 'small-big-small-big' tells the next is big). Replaced with neutral wording.
• G-273: removed parenthetical (the object that ends up floating higher) which directly identifies the answer.
• O-23: removed outcome narrative (Both branches will complete successfully...) and v3's gap-existence hint.
• O-24: removed to find where the chain will break / until the gap prevents further progress.
• O-31: removed algorithm name (using a greedy largest-first strategy) and optimal-order hint (the largest red ball → any red ball). The eating rule (must be smaller-or-equal) is preserved.
• O-45: removed procedural decomposition.

Misc minor cuts (4 tasks)

• G-21: reframed mask motion as a depth-reasoning task.
• O-2: removed Using subtractive color mixing rules,.
• O-16: normalized wording.
• O-22: removed Plan your moves carefully before starting..

Splits

• In-Domain_50: 50 tasks. Training distribution.
• Out-of-Domain_50: 50 tasks. Held out from training.

Excluded from this release

• B-variants (G-21B, G-36B, O-18B, O-19B, G-160B, G-167B, G-218B, G-219B, G-221B): our synthesized depth-reasoning / extremum-flip variants. Not part of upstream VBVR.

Citing VBVR

@misc{vbvr,
  title  = {VBVR: A Very Big Video Reasoning Suite},
  url    = {https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data},
  year   = {2026}
}