datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
vbvr-latent-cache-832x832x33f-t2v-only
VBVR Latent Cache (832×832 × 33f, Wan2.2-TI2V-5B VAE + UMT5-XXL)
Pre-encoded latent cache for the
Video-Reason/VBVR-Dataset
geometric / logical reasoning video corpus, prepared for Equilibrium Matching
(EqM) post-training of Wan-AI/Wan2.2-TI2V-5B-Diffusers on AWS Trainium2.
This is a working cache, not a primary dataset. It exists to skip the
~5 s/sample VAE+T5 encode cost during training. The original videos +
prompts live in the upstream VBVR-Dataset repo.
Source →… See the full description on the dataset page: https://huggingface.co/datasets/Central-Cat/vbvr-latent-cache-832x832x33f-t2v-only.VBVR-Bench-Data
VBVR: A Very Big Video Reasoning Suite
Overview
Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture,
enabling intuitive reasoning over motion, interaction, and causality. Rapid progress in video models has focused primarily on visual quality.
Systematically studying video reasoning and its scaling behavior suffers from a lack of… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Bench-Data.VBVR-Pro-SFT-Image
VBVR-Pro-SFT-Image
The interleaved-image supervised-fine-tuning split of VBVR-Pro: 1.24M programmatically generated reasoning instances across 250 parameterized tasks, one tar.gz per task.
Where VBVR-Pro-SFT-Video asks a model to render the reasoning process as a… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Image.VBVR-Pro-SFT-Video
VBVR-Pro-SFT-Video
The video (I2V) supervised-fine-tuning split of VBVR-Pro: 1.24M programmatically generated reasoning instances across 250 parameterized tasks, one tar.gz per task.
At a glance
Property
Value
Tasks
250
Instances
1,250,000… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Video.VBVR-Pro-RL
VBVR-Pro-RL
The reinforcement-learning split of VBVR-Pro: 50 parameterized tasks × 1,000 instances, held out from the SFT splits, in both a video (TI2V) and an interleaved-image setting.
At a glance
Property
Value
Tasks
50
Instances per… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Pro-RL.VBVR-MultiStep
VBVR-MultiStep
The ~360k-sample programmatic training corpus for long-horizon multi-step image-to-video (I2V) reasoning. Companion to the frozen VBVR-MultiStep-Bench (180-instance evaluation split).
Part of the VBVR (Very Big Video Reasoning Suite) project: https://video-reason.com. See Wang et al., ICML 2026 for the parent suite.
At a glance
Property
Value
Tasks
36 parameterized tasks (Multi-01 … Multi-36)
Reasoning families
Navigation, Planning, CSP… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep.VBVR-Dataset
VBVR-Dataset: Very Big Video Reasoning Training Data
## Overview
VBVR-Dataset is an unprecedentedly large-scale video reasoning training resource, part of the Very Big Video Reasoning (VBVR) Suite. This release contains the training split: 100 curated reasoning task generators with 1,000,000 video clips (10,000 samples per generator), with each sample consisting of a video, start/end frames, a textual reasoning prompt, and… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Dataset.VBVR-Reorganized
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… See the full description on the dataset page: https://huggingface.co/datasets/May-apple/VBVR-Reorganized.VBVR-Reorganized-Image
VBVR-Reorganized-Image
Image-mode derivative of 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… See the full description on the dataset page: https://huggingface.co/datasets/May-apple/VBVR-Reorganized-Image.vbvrpro_sampler_trajectories-data
VBVR-Pro sampler trajectory media
This media archive backs the interactive
pufanyi/vbvrpro_sampler_trajectories
Space.
It contains 12 matched evaluation cells:
DiffSynth step-35500 baseline and DanceGRPO checkpoint 2200
Flow-CPS noise 0.1, 0.3, 0.7, and 0.9
deterministic FlowMatch Euler ODE and UniPC ODE
500 samples per cell across 100 VBVR-Pro tasks
The deployment is split across three public media repositories so each Git-backed
Dataset remains below Hugging Face's… See the full description on the dataset page: https://huggingface.co/datasets/pufanyi/vbvrpro_sampler_trajectories-data.VBVR-Pro-Bench
VBVR-Pro-Bench
The frozen evaluation split of VBVR-Pro. 100 parameterized reasoning tasks, 5 held-out instances each, in both a video (I2V) and an interleaved-image setting.
At a glance
Property
Value
Tasks
100 (In-Domain_50 50 +… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-Pro-Bench.vbvrpro_sampler_trajectories-2200-steps
VBVR-Pro sampler trajectory media
This media archive backs the interactive
pufanyi/vbvrpro_sampler_trajectories
Space.
It contains 12 matched evaluation cells:
DiffSynth step-35500 baseline and DanceGRPO checkpoint 2200
Flow-CPS noise 0.1, 0.3, 0.7, and 0.9
deterministic FlowMatch Euler ODE and UniPC ODE
500 samples per cell across 100 VBVR-Pro tasks
The deployment is split across three public media repositories so each Git-backed
Dataset remains below Hugging Face's… See the full description on the dataset page: https://huggingface.co/datasets/pufanyi/vbvrpro_sampler_trajectories-2200-steps.vbvr-thinkgen
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 v7 prompts used by the ThinkGen video-reasoning experiments.
The snapshot combines:
the prompt revisions from 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… See the full description on the dataset page: https://huggingface.co/datasets/xvreason/vbvr-thinkgen.VBVR-MultiStep-Bench
VBVR-MultiStep-Bench
The frozen 180-instance public evaluation split released alongside the VBVR-MultiStep training corpus. Designed for long-horizon multi-step image-to-video (I2V) reasoning evaluation.
This dataset is part of the VBVR (Very Big Video Reasoning Suite) project. See the parent suite at https://video-reason.com and the suite paper VBVR: A Very Big Video Reasoning Suite (Wang et al., ICML 2026).
At a glance
Property
Value
Tasks
36… See the full description on the dataset page: https://huggingface.co/datasets/Video-Reason/VBVR-MultiStep-Bench.vbvrpro_sampler_trajectories-baseline-steps
VBVR-Pro sampler trajectory media
This media archive backs the interactive
pufanyi/vbvrpro_sampler_trajectories
Space.
It contains 12 matched evaluation cells:
DiffSynth step-35500 baseline and DanceGRPO checkpoint 2200
Flow-CPS noise 0.1, 0.3, 0.7, and 0.9
deterministic FlowMatch Euler ODE and UniPC ODE
500 samples per cell across 100 VBVR-Pro tasks
The deployment is split across three public media repositories so each Git-backed
Dataset remains below Hugging Face's… See the full description on the dataset page: https://huggingface.co/datasets/pufanyi/vbvrpro_sampler_trajectories-baseline-steps.vbvr-ball-bounce-tokencl-trajectory-overlays
VBVR Ball Bounce TokenCL Trajectory Overlay Results
This dataset stores trajectory overlay comparison images for the VBVR ball bounce run:
vbvr_ball_bounce_seed123_batch000001_headaware_step12500_requested_20260613.
Contents
metadata.csv: one row per sample with the winner label and metric columns.
gallery_manifest.json: manifest used by the companion Static HTML Space.
wo_vs_withTokenCL_gt_distance.csv: original per-sample metric table.… See the full description on the dataset page: https://huggingface.co/datasets/zoeloopy/vbvr-ball-bounce-tokencl-trajectory-overlays.vbvr-thinkgen
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 v7 prompts used by the ThinkGen video-reasoning experiments.
The snapshot combines:
the prompt revisions from 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… See the full description on the dataset page: https://huggingface.co/datasets/Kkuntal990/vbvr-thinkgen.VBVR-Bench-Data
VBVR: A Very Big Video Reasoning Suite
Overview
Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture,
enabling intuitive reasoning over motion, interaction, and causality. Rapid progress in video models has focused primarily on visual quality.
Systematically studying video reasoning and its scaling behavior suffers from a lack of video… See the full description on the dataset page: https://huggingface.co/datasets/abs794/VBVR-Bench-Data.vbvr-rl-eval-resultsVBVR-Bench
VBVR-Bench
Re-hosted copy of Video-Reason/VBVR-Bench-Data,
converted to standard HuggingFace parquet format.
Splits
in_domain: 50 tasks x 5 samples = 250 entries (tasks overlap with the VBVR training set).
out_of_domain: 50 tasks x 5 samples = 250 entries (held-out reasoning tasks).
Schema
field
type
notes
task_name
string
e.g. G-13_grid_number_sequence_data-generator
video_idx
string
zero-padded sample id (00000..00004)
domain
string… See the full description on the dataset page: https://huggingface.co/datasets/pufanyi/VBVR-Bench.VBVR-G-13_latentsVBVR-G-15_latentsVBVR-Test-Bench
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… See the full description on the dataset page: https://huggingface.co/datasets/May-apple/VBVR-Test-Bench.vbvr-benchmark-metric-resultsvbvrpro_output-data
VBVR-Pro strict-sweep scored videos
This media archive backs the interactive
pufanyi/vbvrpro_output
Space.
It contains 10,500 scorer-input MP4 files:
20 DanceGRPO results: 5 checkpoints × 4 sampling modes × 500 samples
1 SFT epoch-1 baseline: 500 samples
100 EvalKit-supported tasks, with 5 samples per task and run
1024×1024, 161-frame H.264 videos encoded at 33 FPS
Paths follow:
videos/{run_id}/{domain_folder}/{task_name}/{video_idx}.mp4
The Space contains the task catalog… See the full description on the dataset page: https://huggingface.co/datasets/pufanyi/vbvrpro_output-data.vbvr-grid-through-seed123-failure ---
license: other
language:
- en
pretty_name: VBVR Grid Through Seed123 Failure Videos
size_categories:
- n<1K
tags:
- vbvr
- video
- failure-analysis
- tokencl
---
# VBVR Grid Through Seed123 Failure Videos
This dataset stores MP4 failure cases from
`vbvr_grid_through_seed123_failure`.
## Contents
- `metadata.csv`: one row per video with method, failure type, path, and video metadata.
- `gallery_manifest.json`:… See the full description on the dataset page: https://huggingface.co/datasets/zoeloopy/vbvr-grid-through-seed123-failure.VBVrbvSrVBVR-G-3_latentsVBVR-G-5_latentsVBVR-G-8_latents