Datasetpublic

ShushengYang/minimal_video_pairs-lmms

minimal_video_pairs-lmms Self-contained re-host of facebook/minimal_video_pairs (MVP, the V-JEPA 2 Minimal Video Pairs benchmark) for use with lmms-eval. Unlike the upstream dataset (metadata only; videos hosted elsewhere and fetched by a Makefile), this repo bundles the referenced video clips as per-source .zip files at the repo root. The lmms-eval mvp_sc task downloads this repo and extracts the zips into <HF_HOME>/mvp_sc/<source>/<file>.mp4, matching each metadata row's… See the full description on the dataset page: https://huggingface.co/datasets/ShushengYang/minimal_video_pairs-lmms.

sourceHugging Faceupdated 3mo agoView on Hugging Face

0likes57downloads

Dataset Card

minimalvideopairs-lmms

Self-contained re-host of `facebook/minimal_video_pairs` (MVP, the V-JEPA 2 Minimal Video Pairs benchmark) for use with `lmms-eval`.

Unlike the upstream dataset (metadata only; videos hosted elsewhere and fetched by a Makefile), this repo bundles the referenced video clips as per-source .zip files at the repo root. The lmms-eval mvp_sc task downloads this repo and extracts the zips into <HF_HOME>/mvp_sc/<source>/<file>.mp4, matching each metadata row's video_path. No MVP_VIDEO_ROOT is needed.

Configs / splits

Four configs, each with a mini and a full split: human_object_interactions, robot_object_interactions, intuitive_physics, temporal_reasoning.

python

from datasets import load_dataset
ds = load_dataset("ShushengYang/minimal_video_pairs-lmms", "intuitive_physics", split="mini")

Metadata fields: video_id, video_path, question, answer, candidates, source, question_type.

Videos

Per-source zips (videos_<source>.zip) store arcnames <source>/<file>.mp4, e.g. videos_pt.zip -> pt/video_7186.mp4. lmms-eval's dataset_kwargs: {video: True, cache_dir: mvp_sc} mechanism extracts them.

License & attribution

Benchmark and source media licensing follow the upstream facebook/minimal_video_pairs dataset and its constituent sources (CLEVRER, SSv2, IntPhys, InfLevel, GRASP, Language-Table, STAR, Vinoground, Perception Test). See the upstream dataset card for details. This repo is a packaging convenience for evaluation only.