README.md

metadata

language:
  - en
license: apache-2.0
task_categories:
  - visual-question-answering
  - video-classification
tags:
  - video
  - reasoning
  - benchmark
  - i2v
pretty_name: VBVR-Bench
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: in_domain
        path: data/in_domain-*
      - split: out_of_domain
        path: data/out_of_domain-*

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	duplicates split name; convenient for filtering
prompt	string	task description fed to the I2V model
first_frame	Image (PNG)	I2V condition frame
final_frame	Image (PNG)	expected final frame
ground_truth_video	binary (MP4)	reference video — decode with decord / PyAV

Quick load

from datasets import load_dataset

ds = load_dataset("pufanyi/VBVR-Bench", split="in_domain")
sample = ds[0]
sample["first_frame"]          # PIL.Image
sample["prompt"]               # str
sample["ground_truth_video"]   # raw MP4 bytes

# Decode the video with decord
import decord, io
vr = decord.VideoReader(io.BytesIO(sample["ground_truth_video"]))

Links

• Upstream dataset: Video-Reason/VBVR-Bench-Data
• Evaluation kit: Video-Reason/VBVR-EvalKit
• Project page: video-reason.com