README.md

---
license: apache-2.0
task_categories:
  - video-classification
  - text-to-video
  - image-to-video
language:
  - en
size_categories:
  - 1M<n<10M
tags:
  - reasoning
  - video-generation
  - benchmark
  - vbvr
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
      - split: train_pr_rewritten
        path: parquet/train_pr_rewritten.parquet
      - split: train_pr_unchanged
        path: parquet/train_pr_unchanged.parquet
---

# VBVR-Reorganized

Reorganized + prompt-cleaned + paired-variant-augmented version of VBVR
(Video-Based Visual Reasoning), prepared for **video-generation training**.
The dataset partitions every task into **Pure_Reasoning** (PR) vs
**Instruction_Following** (IF), rewrites Pure_Reasoning prompts to remove
leak phrases, and adds 4 *paired-variant* generators (G-21B/G-36B/O-18B/O-19B)
that share a first frame with their forward counterpart but require the
model to generate a different ground-truth video purely based on a prompt
cue.

The whole tree is built from hardlinks into the original VBVR dataset, so
zero extra disk is consumed for media files.

## Layout

```
VBVR-Reorganized/
├── train/
│   ├── Pure_Reasoning/         (51 generators, 510,000 samples)
│   └── Instruction_Following/  (53 generators, 530,000 samples)
└── test/
    ├── In-Domain_50/
    │   ├── Pure_Reasoning/         (34 generators, 170 samples)
    │   └── Instruction_Following/  (19 generators,  95 samples)
    └── Out-of-Domain_50/
        ├── Pure_Reasoning/         (13 generators,  65 samples)
        └── Instruction_Following/  (42 generators, 210 samples)
```

Sample dir layouts within a generator:
- **train**:  `<gen>/<task>_task/<task>_<8digit>/`  (10,000 samples / generator)
- **test/**: `<gen>/<sample-id>/`  (5 samples / generator)

Each sample directory contains:
- `first_frame.png`, `final_frame.png` — visual context (input + target final state)
- `ground_truth.mp4` — the video the generator should produce
- `metadata.json` — generator-emitted task parameters (e.g. ball trajectory points,
  reward positions, …) for reproducibility / programmatic eval
- `prompt.txt` — the prompt to use as text input
- `prompt_original.txt` — present only for Pure_Reasoning samples on a rule-based-rewrite task; preserves the unmodified original prompt for traceability

## Counts

| Split                  | Class                  | Generators | Samples   |
|------------------------|------------------------|-----------:|----------:|
| train                  | Pure_Reasoning         |         51 |   510,000 |
| train                  | Instruction_Following  |         53 |   530,000 |
| test/In-Domain_50      | Pure_Reasoning         |         34 |       170 |
| test/In-Domain_50      | Instruction_Following  |         19 |        95 |
| test/Out-of-Domain_50  | Pure_Reasoning         |         13 |        65 |
| test/Out-of-Domain_50  | Instruction_Following  |         42 |       210 |
| **TOTAL**              |                        |    **212** | **1,040,540** |

**Note on generator-vs-instance count**: there are **159 unique generator
identities** but each can appear in multiple splits (e.g. G-13 lives in
`train/PR` AND `test/In-Domain_50/PR`).

## Pure_Reasoning subset

The **64 Pure_Reasoning generators** (counted by unique identity) split as:

- **12 already-clean tasks** (no rewrite rule fires): G-43, G-51, G-132,
  G-193, O-8, O-22, O-44, O-47, O-52, O-53, O-56, O-85.
- **47 rule-based-rewrite tasks** organized by family:
  - **A_GRID_GRAPH_SEARCH** (15): G-12/13/14/15/16/17/18, G-31/32, G-41,
    G-44/45/46/47, O-39 — strip "step by step" choreography
  - **B_PATTERN_SEQUENCE** (5): G-131/133/134/135, O-45 — strip pattern leaks
  - **C_ANALOGY** (7): O-7/9/10/11/12/13/14 — strip transformation scripts
    and answer leaks
  - **D_PHYSICS** (13): G-35, G-48, G-273, O-1/2/15/16/17/18/19/62/75/87 —
    strip law names + answer-leaking sentences (parameters preserved)
  - **E_OUTCOME_NARRATIVE** (4): O-23/24/29/31 — strip branch-result and
    stop-point spoilers
  - **F_CANDIDATE_CHECK** (1): O-21 — strip candidate-checking choreography
- **5 promotion tasks** (originally IF, promoted with full rewrites):
  G-21, G-36, O-43, plus 2 paired-variant inverses below.

### Paired training-variant generators (4 unique tasks, 8 generators total)

Same `first_frame.png` as the forward task, but the prompt cues a different
ground-truth video. Designed to teach the video model to actually *read* the
prompt rather than memorise the visual.

| Pair                                              | Mechanic difference                                                              |
|---------------------------------------------------|----------------------------------------------------------------------------------|
| `G-21` ↔ `G-21B_multiple_occlusions_vertical_behind`     | Mask passes **in front of** vs **behind** the objects (depth ordering)        |
| `G-36` ↔ `G-36B_multiple_occlusions_horizontal_behind`   | Same depth flip, horizontal direction                                          |
| `O-18` ↔ `O-18B_glass_refraction_inverse`                | Forward Snell's vs **inverse** Snell's (given in-glass ray, predict incidence) |
| `O-19` ↔ `O-19B_mirror_reflection_inverse`               | Forward reflection vs **inverse** reflection (given reflected ray, predict incidence) |

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

Five additional `*B` variants exist 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
the OOD test split (25 samples total). Counted in the OOD IF total
above; this section just makes them discoverable by name.

## Rewrite stats

- Pure_Reasoning samples covered by rule-based rewriting: **400,180** (47 PR_REWRITE
  tasks × all of their samples in any split, plus the 4 paired variants)
- Prompts actually rewritten by rules: **~320,000** (~80%)
- Prompts left verbatim because no rule fired (already clean): **~80,000** (~20%) —
  these are still PR tasks but their specific paraphrase didn't trip a rule
- Instruction_Following samples (no rewriting): **620,330** — no `prompt_original.txt`

### Inspection splits: `train_pr_rewritten` / `train_pr_unchanged`

Two small viewer-friendly splits (carved out of `train/Pure_Reasoning`) make
the rewrite behavior browsable without loading the full 1M training set:

| Split                  | Rows | Per task        | Selection criterion                                                                 |
|------------------------|-----:|-----------------|-------------------------------------------------------------------------------------|
| `train_pr_rewritten`   |   62 | 2/task          | `prompt != prompt_original` (rule fired and changed the prompt)                     |
| `train_pr_unchanged`   |   46 | 2/task          | `prompt_original` missing OR `prompt == prompt_original` (already-clean or no-op)   |

A given task generator may contribute to one bucket, the other, or both —
see [`TRAIN_PR_REWRITE_CLASSIFICATION.md`](TRAIN_PR_REWRITE_CLASSIFICATION.md) for the per-task breakdown.
Among the 51 train PR tasks: 28 are rewritten-only, 20 are unchanged-only, and 3
(O-10, O-16, O-17) carry multiple paraphrase templates and contribute to *both*
splits (verified by full 10,000-sample scan; per-task rewritten rate is 97.8% /
49.8% / 24.7% respectively).

## How prompt rewrites work

`prompt.txt` is the version to feed to the model. For PR tasks where the
rule engine identified a leak phrase, the rewrite removes it while
preserving setup info needed to make the task solvable. The `prompt_original.txt`
sibling preserves the unrewritten source for traceability.

### `prompt_original` semantics (important)

`prompt_original` field **presence** signals "this sample's task was processed
by the rule-based rewrite engine" — *not* "this sample's prompt was modified".
Whether the prompt was actually changed is determined by comparing strings:

| State                                              | `prompt_original`     | `prompt == prompt_original` |
|----------------------------------------------------|------------------------|------------------------------|
| Instruction_Following task                          | `null`                 | (n/a)                        |
| PR already-clean task (12 tasks: G-43, O-22, …)     | `null`                 | (n/a)                        |
| PR rewrite task, rule fired                         | not null               | **False** — prompt was changed |
| PR rewrite task, no rule fired (already-clean text) | not null               | **True** — prompt unchanged    |

So the field combines two signals: presence ↔ "this task ran the rewrite
engine"; equality ↔ "no rule actually fired on this paraphrase". To filter
to *actually-modified* samples:

```python
modified = ds.filter(lambda x: x["prompt_original"] is not None
                                and x["prompt_original"] != x["prompt"])
# ~320,000 samples (80% of PR-rewrite-task samples; 20% had no rule fire)
```

Family-level rules + task-specific overrides live in
`scripts/vbvr_reorg/rules.py` (in the source repo). See
[PROMPT_REWRITES.md](../VBVR-Reorganized-Samples/PROMPT_REWRITES.md) (in the
`VBVR-Reorganized-Samples/` subset) for a representative before→after diff
of every rewritten task.

## How to use this dataset

> **The HF repo ships parquet shards, not the directory tree above.** The
> `Layout` section describes the *conceptual* organization. On Hub, every
> sample is one parquet row with media bytes (`first_frame.png`,
> `final_frame.png`, `ground_truth.mp4`) embedded inline. You have two
> options: read the parquet directly (Path A), or materialize the parquet
> back into the directory layout (Path B).

### Parquet row schema (all splits)

Every parquet shard under `parquet/` shares the same row schema:

| Column              | Type                                        | Notes                                                          |
|---------------------|---------------------------------------------|----------------------------------------------------------------|
| `class`             | string                                      | `Pure_Reasoning` or `Instruction_Following`                    |
| `task`              | string                                      | e.g. `G-13_grid_number_sequence_data-generator`                |
| `split`             | string                                      | e.g. `train`, `test/In-Domain_50`, `train_pr_rewritten`        |
| `sample_id`         | string                                      | per-task relative path (for train rows includes the wrapper)   |
| `prompt`            | string                                      | the prompt to feed to the model                                 |
| `prompt_original`   | string \| null                              | only for PR-rewrite tasks (see "How prompt rewrites work")     |
| `first_frame`       | `{bytes: binary, path: string}` (HF Image)  | PNG bytes of the input frame                                   |
| `final_frame`       | `{bytes: binary, path: string}` (HF Image)  | PNG bytes of the target final frame                            |
| `ground_truth_video`| `{bytes: binary, path: string}` (HF Video)  | MP4 bytes of the target video                                  |
| `metadata_json`     | string                                      | raw text of the per-sample `metadata.json`                      |

Available splits / parquet shards:

| Logical split           | Parquet path                            | Total size |
|-------------------------|-----------------------------------------|-----------:|
| `train`                 | `parquet/train__<task>.parquet` (107 shards, one per task) | ~370 GB |
| `test_in_domain`        | `parquet/test_in_domain.parquet`        | ~90 MB    |
| `test_out_of_domain`    | `parquet/test_out_of_domain.parquet`    | ~120 MB   |
| `train_pr_rewritten`    | `parquet/train_pr_rewritten.parquet`    | ~19 MB    |
| `train_pr_unchanged`    | `parquet/train_pr_unchanged.parquet`    | ~14 MB    |

**Important — picking which train shards to download**: train is sharded
per task (51 Pure_Reasoning shards + 56 Instruction_Following shards =
107 shards, ~370 GB total). Most users do **not** want all 370 GB. See
[**TASK_INVENTORY.md**](TASK_INVENTORY.md) for the full per-task table
(shard name ↔ full task id ↔ class ↔ size). The same data is also
shipped as machine-readable `TASK_INVENTORY.json` so you can script
selective downloads. **The shard filename alone does NOT tell you
whether a task is PR or IF** — use the inventory.

### Path A — read parquet rows directly (recommended)

The HF `datasets` library auto-decodes `first_frame` / `final_frame` into
`PIL.Image` and `ground_truth_video` into a `decord`/`VideoReader`-style
object:

```python
from datasets import load_dataset

# Load any single split. Shards under parquet/train__*.parquet are merged
# into one logical "train" split via the YAML config above.
ds = load_dataset("May-apple/VBVR-Reorganized", split="test_in_domain")
# other splits: "train", "test_out_of_domain", "train_pr_rewritten", "train_pr_unchanged"

row = ds[0]
print(row["prompt"])            # str
print(row["class"], row["task"])
row["first_frame"].save("first.png")   # PIL.Image
row["final_frame"].save("final.png")
# ground_truth_video is a `datasets.Video` object; underlying bytes are in
# row["ground_truth_video"]["bytes"] if you bypass auto-decoding (see below).
```

If you only want the **prompts / metadata** and don't need to decode the
heavy media (much faster, no PIL/decord needed), bypass HF's feature
decoding and read the raw parquet with `pyarrow` / `pandas`:

```python
import pyarrow.parquet as pq

t = pq.read_table("parquet/test_in_domain.parquet",
                  columns=["class", "task", "split", "sample_id",
                           "prompt", "prompt_original", "metadata_json"])
df = t.to_pandas()
print(df.head())
```

To pull the **raw bytes** of one media field for a single row:

```python
import pyarrow.parquet as pq

pf = pq.ParquetFile("parquet/test_in_domain.parquet")
row = pf.read_row_group(0).to_pylist()[0]
open("first_frame.png", "wb").write(row["first_frame"]["bytes"])
open("final_frame.png", "wb").write(row["final_frame"]["bytes"])
open("ground_truth.mp4", "wb").write(row["ground_truth_video"]["bytes"])
```

To load only **one train task** (without pulling the whole 1M-row
training set), just load that task's parquet shard by filename:

```python
from datasets import load_dataset
ds = load_dataset("parquet",
                  data_files="parquet/train__G-13_grid_number_sequence.parquet",
                  split="train")
```

#### Filtering PR vs IF in Path A

Every row carries a `class` column with value `"Pure_Reasoning"` or
`"Instruction_Following"`. Two ways to use it:

```python
# (1) Load all 107 shards, filter in code (requires downloading all ~370 GB)
from datasets import load_dataset
ds = load_dataset("May-apple/VBVR-Reorganized", split="train")
pr_only = ds.filter(lambda x: x["class"] == "Pure_Reasoning")
if_only = ds.filter(lambda x: x["class"] == "Instruction_Following")

# (2) Load only the PR shards (downloads only ~195 GB) — use TASK_INVENTORY.json:
import json
from datasets import load_dataset
inv = json.load(open("TASK_INVENTORY.json"))  # ship from this repo
pr_shards = [f"parquet/train__{e['shard']}.parquet" for e in inv["Pure_Reasoning"]]
ds_pr = load_dataset("parquet", data_files=pr_shards, split="train")

# Same for IF:
if_shards = [f"parquet/train__{e['shard']}.parquet" for e in inv["Instruction_Following"]]
ds_if = load_dataset("parquet", data_files=if_shards, split="train")
```

`TASK_INVENTORY.json` is shipped at the repo root specifically so you
can do this kind of selective load without enumerating shards by hand.
For the canonical authoritative source on which task is PR vs IF, see
[TASK_INVENTORY.md](TASK_INVENTORY.md).

### Path B — materialize parquet → on-disk directory layout

If your existing pipeline reads `first_frame.png` / `ground_truth.mp4` /
`prompt.txt` from a directory tree, this repo ships an extractor script
(`extract_parquet_to_dir.py`, in the repo root) that streams a parquet
shard and writes one directory per row, reproducing exactly the "Layout"
tree at the top of this card:

Only dependency: `pyarrow` (`pip install pyarrow`).

#### B.1 — Standard workflow: download everything, extract everything

```bash
# (a) Pull the whole repo (~370 GB parquet + tiny test/inspection splits + helpers).
hf download May-apple/VBVR-Reorganized --repo-type dataset \
    --local-dir ./VBVR-Reorganized
cd ./VBVR-Reorganized

# (b) Materialize every parquet under parquet/ into ./extracted/.
#     Streams one row-group at a time, safe on the multi-GB train shards.
python3 extract_parquet_to_dir.py \
    --parquet parquet/ \
    --out ./extracted
```

After this, `./extracted/` is the exact tree shown in the "Layout"
section at the top of this card:

```
extracted/
├── train/
│   ├── Pure_Reasoning/         (51 generators, 510,000 samples)
│   └── Instruction_Following/  (56 generators, 560,000 samples)
├── test/
│   ├── In-Domain_50/...
│   └── Out-of-Domain_50/...
├── train_pr_rewritten/...      (inspection split, 62 samples)
└── train_pr_unchanged/...      (inspection split, 46 samples)
```

> **Disk note**: extracted media files are the same bytes that already
> live in the parquet (no recompression), so the extracted tree is
> roughly the same size as the parquet (~370 GB). Peak disk usage during
> extraction is ~2× that since both copies exist. If disk is tight,
> extract per-shard and delete the source parquet after each: see B.2.

#### B.2 — Subset variants

```bash
# Just the test splits (~210 MB combined, fast):
hf download May-apple/VBVR-Reorganized --repo-type dataset \
    --include 'parquet/test_*.parquet' 'extract_parquet_to_dir.py' \
    --local-dir ./VBVR-Reorganized
cd ./VBVR-Reorganized
python3 extract_parquet_to_dir.py --parquet parquet/ --out ./extracted

# A single task (e.g. G-13, ~5.96 GB) — see TASK_INVENTORY.md for shard names:
hf download May-apple/VBVR-Reorganized --repo-type dataset \
    --include 'parquet/train__G-13_grid_number_sequence.parquet' \
              'extract_parquet_to_dir.py' \
    --local-dir ./VBVR-Reorganized
cd ./VBVR-Reorganized
python3 extract_parquet_to_dir.py \
    --parquet parquet/train__G-13_grid_number_sequence.parquet \
    --out ./extracted

# Per-shard extract-then-delete (avoid peak 2× disk for the full pull):
cd ./VBVR-Reorganized
for f in parquet/train__*.parquet; do
    python3 extract_parquet_to_dir.py --parquet "$f" --out ./extracted
    rm "$f"
done
```

For class-level downloads (e.g. *all* PR train shards = ~195 GB), see
the "Selective download recipes" section in
[TASK_INVENTORY.md](TASK_INVENTORY.md).

Output tree:

```
VBVR-extracted/
└── <split>/<class>/<task>/<sample_id>/
    ├── first_frame.png
    ├── final_frame.png
    ├── ground_truth.mp4
    ├── prompt.txt
    ├── prompt_original.txt    (only for PR-rewrite tasks)
    └── metadata.json
```

The script reads one parquet row-group at a time, so it works on the
multi-GB train shards without loading the whole file into memory. The
script body is short (<150 lines, pure stdlib + `pyarrow`) — copy it
directly if you don't have access to the source repo.

### Local browsing (small viewable subset)

For a 3-samples-per-generator view (612 samples, ~2GB), see the
`VBVR-Reorganized-Samples/` companion repo / subset.

## Provenance

This is a **derivative** of the original VBVR dataset. The original dataset
([Video-Reason/VBVR-Bench-Data](https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data) and the
1M-sample train extract) is preserved untouched at:

- `<source>/VBVR-Dataset-extracted/train/`
- `<source>/VBVR-Bench-Data/{In-Domain_50,Out-of-Domain_50}/`

This dataset is built from those via hardlinks (zero copy of media); only
`prompt.txt` files are physically rewritten.

## Reproduce

The full reorganization + rewrite pipeline is in `scripts/vbvr_reorg/` of the
source repo (not bundled here):

- `classification.py` — 64 PR + 91 IF + family labels (the "what is what")
- `rules.py` — rewrite rules (task-specific + family-level)
- `build_and_rewrite.py` — single-pass: hardlink media + apply rewrites
- `reapply_rules.py` — refresh rewrites in place when `rules.py` changes
- 9 paired-variant generators forked from `O-2`/`O-18`/`O-19`/`G-21`/`G-36`/etc.

## Citation

```bibtex
@dataset{vbvr_reorganized_2026,
  title  = {VBVR-Reorganized: Reasoning-vs-Instruction Partition with Prompt-Cleaning and Paired-Variant Augmentation for Video Generation},
  author = {Video-Reason},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Video-Reason/VBVR-Reorganized},
}
```

## License

Inherits the license of the underlying VBVR dataset (Apache-2.0 unless
otherwise specified by the original publishers).