May-apple/VBVR-Reorganized-Image

---
license: apache-2.0
task_categories:
- text-to-image
- image-to-image
language:
- en
size_categories:
- 100K<n<1M
tags:
- reasoning
- image-generation
- benchmark
- vbvr
- image-mode
configs:
- config_name: default
data_files:
- split: train
path: parquet/train__*.parquet
- split: train_samples
path: parquet/train_samples.parquet
- split: test_in_domain
path: parquet/test_in_domain__*.parquet
- split: test_out_of_domain
path: parquet/test_out_of_domain__*.parquet
---

# VBVR-Reorganized-Image

Image-mode derivative of [VBVR-Reorganized](https://huggingface.co/datasets/May-apple/VBVR-Reorganized).
Each sample is a triple `(first_frame.png, prompt.txt, final_frame.png)`:
the model takes `first_frame + prompt` as input and should output an
image that matches `final_frame`. **No video** in this version — purely
single-image-input, single-image-output.

## Layout

```
VBVR-Reorganized-Image/
├── train/
│ ├── Pure_Reasoning/ (48 generators, 480,000 samples)
│ └── Instruction_Following/ (48 generators, 480,000 samples)
└── test/
├── In-Domain_50/
│ ├── Pure_Reasoning/ (31 generators, 155 samples)
│ └── Instruction_Following/ (17 generators, 85 samples)
└── Out-of-Domain_50/
├── Pure_Reasoning/ (11 generators, 55 samples)
└── Instruction_Following/ (42 generators, 210 samples)
```

Each sample directory contains exactly three files:
- `first_frame.png` — visual input
- `final_frame.png` — image-mode ground truth (target output)
- `prompt.txt` — text input (already cleaned for image-mode)

## Counts

| Split | Class | Generators | Samples |
|------------------------|------------------------|-----------:|----------:|
| train | Pure_Reasoning | 48 | 480,000 |
| train | Instruction_Following | 48 | 480,000 |
| test/In-Domain_50 | Pure_Reasoning | 31 | 155 |
| test/In-Domain_50 | Instruction_Following | 17 | 85 |
| test/Out-of-Domain_50 | Pure_Reasoning | 11 | 55 |
| test/Out-of-Domain_50 | Instruction_Following | 42 | 210 |
| **TOTAL** | | **197** | **960,505** |

## How this differs from the video-mode parent

- **No `ground_truth.mp4`** — image-mode tasks have a single static answer image instead of a video.
- **No `metadata.json`** — task parameters not exposed at row level (still recoverable from the parent video repo if needed).
- **Only one prompt per sample** (`prompt.txt`); `prompt_original.txt` is dropped to keep rows lean.
- **CLASS_3 tasks dropped** — 10 task types (e.g. `O-22_construction_stack`,
`G-39_attention_shift_different`, `O-32_rolling_ball`, `O-44_rotation_puzzle`,
`O-47_sliding_puzzle`, `O-52_traffic_light`, `O-62_gravity_physics`,
`G-11_handle_object_reappearance`, `G-22_attention_shift_same`,
`G-33_visual_jenga`) are temporal-by-nature tasks whose single-image
version carries no reasoning signal. They are excluded entirely.

## Image-mode classes

The 197 task-split slots fall into two construction classes:

| Class | Count | `final_frame.png` source | Prompt rewriter |
|-------|------:|--------------------------|-----------------|
| **CLASS_1** | 171 | Copied verbatim from the video-mode last frame | Light cleanup of process language ("step by step", "render the X", etc.) via `prompt_rewriter.py` / `train_prompt_rules.py` |
| **CLASS_2** | 26 | **Re-rendered** from `metadata.json` by a per-task painter (orange path cells for grid/maze tasks, red trajectory polylines for bouncing balls, numbered labels on fallen dominoes, ...) | Original prompt + appended task-specific image-mode output instruction |

CLASS_2 examples:
- Grid/maze (G-12 to G-18, G-31, G-32, G-41, G-44 to G-47, O-39): orange path overlay
- Physics (G-35, G-48, O-15): red trajectory polyline
- Domino (O-23, O-24): numeric labels on fallen pieces
- Occlusion (G-21, G-36): mask redefined to stop at object midline
- Other: O-29, O-31, O-34

## Pure_Reasoning prompt cleanup

For Pure_Reasoning tasks, prompts are stripped of reasoning leaks beyond
the standard image-mode cleanup. The full leak-removal pipeline runs:
`rules.py` (family-level + task-specific rules from the video-mode dataset)
+ `rules_image.py` (image-mode-specific paraphrase handlers).

Examples of stripped leaks:
- O-23 (E_OUTCOME_NARRATIVE): drop the 4-sentence outcome narration
("trunk falls first, then splits into Branch A...")
- O-12 / O-11 / O-13 / O-14 (C_ANALOGY): drop the explicit
"first change its color, then change its size" enumeration
- G-273 (D_PHYSICS): drop the answer-leaking "right container holds the
higher-density liquid" + parenthetical pointer
- O-15 (D_PHYSICS): drop "elastic collision physics (angle of incidence
equals angle of reflection)"
- O-75 (D_PHYSICS): drop the terminal-state spoiler "to a common
equilibrium level across all tubes"
- O-45 (B_PATTERN_SEQUENCE): drop "Observe the cyclic order... Identify
the color cycle..." choreography for both color and arithmetic paraphrases

Constraint phrases that are **kept** (they specify the task, not the
answer): "shortest path", "minimum number of steps", "additive color
mixing", "subtractive color mixing", physics constants (refractive
index, viscous damping coefficient).

## Paired-variant generators (4 unique tasks)

The same image-mode pipeline carries the depth-flip and inverse variants
created in the parent video-mode dataset:

| Variant | Mechanic difference vs forward |
|---------|--------------------------------|
| `G-21B_multiple_occlusions_vertical_behind` | Mask passes **behind** (objects in front) — final_frame: mask gone, objects visible |
| `G-36B_multiple_occlusions_horizontal_behind` | Same depth flip, horizontal direction |
| `O-18B_glass_refraction_inverse` | Given in-glass ray, predict incidence ray |
| `O-19B_mirror_reflection_inverse` | Given reflected ray, predict incidence ray |

These share the same `first_frame.png` as their forward counterpart but
have a different `final_frame.png` and a prompt that distinguishes the
direction. The pair tests whether the model is reading the prompt rather
than memorising the visual.

### Tier-2 extremum-flip variants (5 unique tasks, test-only)

Five additional `*B` variants live in
`test/Out-of-Domain_50/Instruction_Following`, flipping the extremum
criterion of their forward task:

| Variant | Forward | Flip |
|---------|---------|------|
| `G-160B_circle_smallest_numerical_value` | `G-160_circle_largest_numerical_value` | largest → **smallest** |
| `G-167B_select_shortest_polygon_side` | `G-167_select_longest_polygon_side` | longest → **shortest** |
| `G-218B_identify_smallest_angle_in_triangle` | `G-218_identify_largest_angle_in_triangle` | largest → **smallest** |
| `G-219B_select_rightmost_shape` | `G-219_select_leftmost_shape` | leftmost → **rightmost** |
| `G-221B_outline_outermost_square` | `G-221_outline_innermost_square` | innermost → **outermost** |

These are **classified as Instruction_Following, not Pure_Reasoning** —
they're explicit-criterion mark-and-pick tasks (mechanical
perception+comparison), so flipping the criterion only changes which
shape gets marked, not the reasoning structure. Each has 5 samples in
OOD (25 samples total). They are counted in the OOD IF total in the
counts table.

## How to use

```python
from datasets import load_dataset

ds = load_dataset("May-apple/VBVR-Reorganized-Image", split="train")
# Each row: class, task, split, sample_id, prompt, first_frame, final_frame
# first_frame and final_frame are HF Image() — call as .convert("RGB") to
# get a PIL image, or pass directly to your model's preprocessor.
```

Three splits:
- `train` — 960,000 samples
- `test_in_domain` — 240 samples
- `test_out_of_domain` — 265 samples

## Provenance

This is a derivative of the parent video-mode dataset. The image-mode
build pipeline lives in the source repo (`scripts/vbvr_reorg/`):
- `build_image_mode_full.py` — flattens video samples into image-mode samples
- `build_parquet_shards_image.py` — packs into HF parquet shards
- `rules_image.py` — image-mode-specific PR leak rules
- Renderers reused from `VBVR-Bench-Image/regenerator/` and
`VBVR-Train-Image/regenerator/`

## Citation

```bibtex
@dataset{vbvr_reorganized_image_2026,
title = {VBVR-Reorganized-Image: Single-Image Reasoning Benchmark Derived from VBVR},
author = {Video-Reason},
year = {2026},
url = {https://huggingface.co/datasets/May-apple/VBVR-Reorganized-Image},
}
```

## License

Inherits the license of the underlying VBVR dataset (Apache-2.0).