---
license: apache-2.0
language:
- en
pretty_name: VBVR ThinkGen
size_categories:
- 10K<n<100K
task_categories:
- text-to-video
tags:
- video-reasoning
- visual-reasoning
- synthetic
- image-to-video
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.csv
---


# VBVR ThinkGen

VBVR ThinkGen is the current ThinkGen training snapshot derived from the balanced VBVR corpus. It contains **99,958 unique videos across 100 generator families**, paired with the **v8 prompts** used by the ThinkGen video-reasoning experiments.

The snapshot combines:

- the prompt revisions from [`May-apple/VBVR-Reorganized`](https://huggingface.co/datasets/May-apple/VBVR-Reorganized),
- the additive ThinkGen v3-v5 prompt revisions over previously untouched families, and
- the **v6 owner-review pass (2026-09-08)**: 30 families / 30,000 prompts edited on top of v5 — 14 fixes to already-revised families (restoring task-process clauses such as "show the full process step by step", "following the law of reflection", "circle ... with a red circle" that earlier passes had deleted, plus further trims) and 16 first-time trims of families no earlier pass had touched (removing scene details readable from the first frame: object colours/shapes, chart titles and axis labels, the current clock time, direction lists, explanatory glosses).
- the **v7 checklist desk pass (2026-09-10)**: 21 families / 20,380 prompts edited on top of v6 after applying the reasoning-inducing prompt checklist to all 100 families — 5 defect repairs (G-45 stray periods in every row, O-7 "a unknown" colour tokens in 998 rows, O-8 wrong answer-slot wording, G-27 stranded articles, O-1 process clause), 11 restorations of task-process, invariant and stopping-rule clauses that the Reorganized pass had deleted (G-46, G-133, G-134, O-4, O-13, O-14, O-21, O-23, O-24, O-29, O-31), 2 well-posedness additions (G-12 leg convention verified in the generator source; G-51 "The colors repeat in a fixed cycle."), and 3 leak/echo removals (O-12 transformation narrative and opaque colour ids, O-75 numeric level list, O-45 rule-type name).
- the **v8 ground-truth pass (2026-10-06)**: 12 families / 8,651 prompts edited on top of v7 after checking every family's prompt against the per-clip generator parameters that ship in the source dataset's `metadata.parquet` (and, for grid routes, against the ground-truth videos). Each edit corrects a statement the ground truth contradicts or states a rule the ground truth obeys in every clip: G-19 (503 clips said "vertical column" where the video ends in a horizontal row), O-3 (499 clips with 6-8 symbols are laid out in two rows), O-13 (743 clips said "filled" for an outlined shape), O-44 (the "rotate 90 degrees" action was false; the goal is one closed loop), O-7 (restores where the answer shape goes), O-52 (a light that is yellow turns red next and switches when its countdown reaches 0), G-41 (no cell is visited twice), G-18 (horizontal-then-vertical on ties between equally short paths), O-21 (a rotated piece does not match the gap), G-27 (the Total row is not a candidate), G-29 ("week" corrected to "day" in 128 rows) and O-10 (24 rows named the transformation to be inferred).

The video set is matched to the balanced ThinkGen training subset; prompt revisions do not replace or regenerate the source videos.

## Versions

Every prompt snapshot is a git tag on this repo; the archives never change between versions.

| Tag | Snapshot | `metadata.csv` SHA-256 | Note |
|---|---|---|---|
| `v8` (= `main`) | ThinkGen VBVR v8 | `681ab1c37daf1a0b56364f24c93d26d5d3f7854001fcce33d1156e0a44cb0d7d` | current |
| `v7` | ThinkGen VBVR v7 | `11a7a6b8d34bbc28b16e62e3b1128ce9d47a251c6982918428f8e5b511dfe434` | previous |
| `v6` | ThinkGen VBVR v6 | `9049cdaca4fd7d594b763d5b9532148fe8b135be86a94db8a2bebc6262124e8c` | older |
| `v5` | ThinkGen VBVR v5 | `a8f156b0a4c45f7a13b7d7eb4cf640d5e3b8ba3b04e777241d385ce2ab7206ab` | older |

To train on the previous prompts, fetch the metadata at the tag:

```bash
hf download xvreason/vbvr-thinkgen metadata.csv --repo-type dataset --revision v7
```

The v6, v7 and v8 edits are fully specified as per-family regex rule files (`scripts/data/review_edits_v6/`, `scripts/data/review_edits_v7/`, `scripts/data/review_edits_v8/`) and applied by `scripts/data/apply_review_edits_v6.py` in the `Kkuntal990/vbvr-realignment` repository; re-running them on the `v5`, `v6` and `v7` metadata reproduces each file byte-for-byte.

## Files and layout

`metadata.csv` contains:

| Column | Description |
|---|---|
| `video` | Dataset-root-relative path to a `ground_truth.mp4` video |
| `prompt` | Training prompt associated with that video |

Prompts can contain embedded newlines. Read `metadata.csv` with an RFC-compliant CSV parser rather than line-oriented shell tools.

The media are stored as **100 uncompressed family archives** under `archives/`. MP4 is already compressed, so uncompressed tar avoids wasted CPU and permits fast extraction. Each archive preserves the paths referenced by `metadata.csv`, for example:

```text
G-11_handle_object_reappearance_data-generator/
  handle_object_reappearance_task/
    handle_object_reappearance_00000000/
      ground_truth.mp4
```

After downloading a snapshot, extract the archives into the snapshot root:

```bash
for archive in archives/*.tar; do tar -xf "$archive"; done
```

The `video` values in `metadata.csv` then resolve directly from that root.

## Snapshot statistics

- Rows: 99,958
- Unique videos: 99,958
- Generator families: 100
- Video payload: 30,467,788,636 bytes
- Prompt revision: ThinkGen v8
- Families whose prompts differ from the original VBVR text: 69
- Families edited in v6 relative to v5: 30 (30,000 prompts)
- Families edited in v7 relative to v6: 21 (20,380 prompts)
- Families edited in v8 relative to v7: 12 (8,651 prompts)

`dataset_manifest.json` records the metadata hash, snapshot-level counts and the previous snapshot's hash/tag. `archives_manifest.json` records the byte size, SHA-256, and video count for every family archive.

## Loading metadata

```python
from datasets import load_dataset

ds = load_dataset("xvreason/vbvr-thinkgen", data_files="metadata.csv", split="train")
```

Use `snapshot_download` or `hf download` to fetch the complete media snapshot before extraction.

## Provenance and license

The source corpus is [`Video-Reason/VBVR-Dataset`](https://huggingface.co/datasets/Video-Reason/VBVR-Dataset). Reorganized prompts come from [`May-apple/VBVR-Reorganized`](https://huggingface.co/datasets/May-apple/VBVR-Reorganized). Both source repositories are Apache-2.0 licensed; this derivative snapshot is released under the same license.

This dataset contains synthetic benchmark videos. Users should validate prompts and family-specific evaluators for their intended research setting.