shivank21/mvbench-temporal-conflict-subset

MVBench Temporal-Conflict Subset

A curated 75-sample subset of OpenGVLab/MVBench selected for a controlled temporal-conflict benchmark on vision-language models.

The full design and motivation are described in the parent project proposal (Temporal Conflict Resolution in Vision-Language Models). In short: each video here was chosen because its question + wrong-option distractors map directly onto a visually-realizable mid-clip edit — recolor, resize, swap, multiplication, or existence-flip — so that the wrong candidate becomes a valid edit target. The benchmark then probes how a VLM arbitrates between the original and the edited state when answering the original MVBench question.

Composition

Split	Source	n	Conflict mechanism
moving_attribute	CLEVRER (synthetic)	15	Recolor / reshape / re-material the target object mid-clip toward a wrong-option distractor
moving_count	CLEVRER (synthetic)	15	Clone or remove objects mid-clip to shift count toward a wrong-option distractor
object_existence	CLEVRER (synthetic)	15	Add or remove the queried object class mid-clip to flip yes ↔ no
moving_direction	CLEVRER (synthetic)	15	Recolor / reshape the named referent mid-clip; original direction question now has two candidate referents
object_interaction	Charades / STAR (real)	15	Real-world object swap inside the action window — replace the held object with a wrong-option distractor

Total: 75 samples, 75 unique videos, ~98 MB.

Selection criteria

Sampling was uniform random (seed 42) from each MVBench JSON, oversampling to 30 candidates per split. Each candidate was then manually verified by a per-split agent against the following editability test:

1. Atomic distractors. The wrong-option candidates must be visually-realizable as a single attribute edit, not narrative descriptions.
2. Referent specificity. The question must name a single, isolatable object (or class) so the edit lands cleanly.
3. QA invariance. The edit must not break the question's logical pre-conditions (e.g., for moving_count, exclude collision-counting questions where cloning would create unscheduled collisions).
4. Editability of the swap target. For object_interaction, every candidate must be a hand-sized, portable object — answers that are furniture (table, bed, closet, refrigerator) were dropped because such swaps are not coherent for current real-world video edit models.

Candidates that failed were replaced from the reserve until 15 valid samples were collected per split. The editability_notes field on every row records the recommended edit type for that sample.

Files

• data/metadata.parquet — the canonical 75-row table (loadable with datasets.load_dataset).
• metadata.json — same data as JSON for non-parquet workflows.
• videos/*.mp4 — 75 video files, named with the original MVBench filename so they can be cross-referenced against OpenGVLab/MVBench.

Schema

Column	Type	Description
id	string	Unique key, e.g. moving_attribute_00
split	string	Original MVBench split name
source_dataset	string	CLEVRER or Charades/STAR
video	string	Original MVBench video filename (also the basename in videos/)
video_path	string	Path relative to repo root
question	string	Original MVBench question (unchanged)
candidates	list[string]	Original 3-5 multiple-choice options
answer	string	Ground-truth answer in the un-edited video
conflict_mechanism	string	Recommended conflict-edit type for this split
editability_notes	string	Per-sample edit recipe written by the verification agents
orig_mvbench_index	int	Index into the original MVBench split JSON
start, end, accurate_start, accurate_end	float	(object_interaction only) Action-window timestamps in seconds for clipping

Usage

from datasets import load_dataset
ds = load_dataset("shivank21/mvbench-temporal-conflict-subset", split="train")
print(ds[0])
# Video is at: hf://datasets/shivank21/mvbench-temporal-conflict-subset/videos/<filename>

To download a single video:

from huggingface_hub import hf_hub_download
local = hf_hub_download(
    repo_id="shivank21/mvbench-temporal-conflict-subset",
    filename=f"videos/{row['video']}",
    repo_type="dataset",
)

Provenance and licensing

• Questions, candidates, answers, and video selections come from OpenGVLab/MVBench (MIT-licensed). Cite their paper arXiv:2311.17005.
• CLEVRER videos are synthetic clips from Yi et al., 2019 (clevrer.csail.mit.edu). Academic-research use only.
• STAR / Charades videos are real-world clips from Sigurdsson et al., 2016 (prior.allenai.org/projects/charades). Subject to the original Charades license.

This subset is for academic research use only.