Sterzhang/motionarean-bench

MotionArena Bench

A standalone motion-understanding benchmark containing 1,000 test examples, balanced across five question families (200 examples each). Each example contains a multiple-choice question, its ground-truth answer, and 16 chronological 1280 × 720 JPEG frames. Frames are embedded in self-contained Parquet shards as a list of Hugging Face Image features for display in the Dataset Viewer.

This repository contains the test split only. It is a byte-identical extraction of the corrected Motion-v3 test split from Sterzhang/motion-train, with the existing questions, options, answers, IDs, frames, and provenance retained.

Usage

from datasets import load_dataset

test = load_dataset("Sterzhang/motionarean-bench", split="test")
example = test[0]
print(example["question"])
print(example["options"], example["answer"], example["answer_letter"])
example["frames"][0]  # a PIL image

For streaming access:

test = load_dataset("Sterzhang/motionarean-bench", split="test", streaming=True)

Category composition

Category	Test examples	Share
gripper_width	200	20%
gripper_rel_to_obj	200	20%
obj_rel_to_camera	200	20%
obj_rel_to_obj	200	20%
trajectory_shape	200	20%
Total	1,000	100%

Fields

• id, question, options, answer, answer_index, answer_letter: the QA. Answer indices are zero-based; answer letters start at A.
• frames: 16 original RGB JPEG frames in chronological order.
• frame_indices: source-relative indices provided by the export. Original source video frame rate and absolute source time are unavailable.
• category: one of the five question families above.
• split, source_id, image_paths, label_revision: export provenance; split is always test, and frame bytes are embedded in frames.
• generated_source_index, source_index: source-export indices; unused fields are null.
• reported_answer, reported_answer_letter: retained publisher correction provenance where present; otherwise null.

Evaluation

Use all 16 supplied frames in chronological order and score the selected option against answer_letter or the zero-based answer_index. Report accuracy over all 1,000 examples and optionally per-family accuracy over each group of 200. Keep this split out of model fitting and checkpoint selection.

Provenance

benchmark_manifest.json records the source revision, canonical QA checksum, category counts, validation results, and each Parquet shard's SHA-256 checksum. SHA256SUMS provides file checksums. All 1,000 QA records and all 16,000 embedded frame checksums were checked before publication.

• Source release: Sterzhang/motion-train
• Original test export: Sterzhang/motion-test
• MotionLM code: plzdoo/MotionLM
• Underlying DROID project: droid-dataset.github.io

The source MotionLM exports did not declare a license in their dataset cards. This extraction does not assign a new license; consult the upstream sources for applicable terms.