KeepaceAI/MoveBench-Rubrics

MoveBench Rubrics and Answer Keys

An annotation-only companion to KeepaceAI/MoveBench.

This separate dataset contains annotations for the existing 80 public questions: 30 multiple-choice answer keys and 464 rubric criteria across 50 open-ended questions. Join records to the original questions using problem_id. This repository does not repeat question text or choices and does not include open-ended reference answers. The original questions-only dataset remains unchanged.

Original questions · Paper · GitHub

Files and fields

• annotations.jsonl: default HF test split, 80 records.
• annotations.json: the same records as a JSON array.
• problem_id: stable ID matching the original public question.
• answer_key: correct MCQ option labels, as a list; empty for open-ended questions. Multi-select answers are compared as sets.
• rubric: a list of dimension, criterion, description, and score fields; empty for MCQ.

Criterion scores are the source point values, not observed model results. Negative safety markers are preserved as published. This file is not a complete executable scoring protocol. Per the current project rule, a confirmed red-line violation incurs −10 points that cannot be offset by positive criterion scores and affects the overall score.

Usage

from datasets import load_dataset

questions = load_dataset("KeepaceAI/MoveBench", split="test")
annotations = load_dataset("KeepaceAI/MoveBench-Rubrics", split="test")
annotations_by_id = {row["problem_id"]: row for row in annotations}
for question in questions:
    annotation = annotations_by_id[question["problem_id"]]
    # Generate model answers from question inputs only.
    # Use annotation["rubric"] or annotation["answer_key"] for evaluation.

Do not pass answers or rubrics to the student model when generating evaluation responses.

Expert contribution and rubric development

According to the project team's production records, 26 experts contributed to the full 809-question benchmark over 20 days, with 2,912 expert-hours across question drafting, answer writing, review, re-review and inclusion in the benchmark. The cohort comprises 14 internal sports-domain specialists and 12 professors and discipline leaders from Capital University of Physical Education and Sports. These figures refer to the entire project, not solely rubric writing or the 80-question release. They do not imply that all 26 experts reviewed every item.

The team reports 2,059,649 characters of project material, approximately 2,546 characters and 3.6 expert-hours per question. These are project-reported production statistics, not independently audited measurements.

The published methodology combines scenario-specific question and rubric drafting, automated quality checks, and expert review. Criteria articulate assessable requirements such as goal alignment, exercise selection, clarity, and safety. Review considers question clarity, answer/rubric consistency, scientific validity and safety; items may be revised or excluded before inclusion.

Acknowledged internal experts

王鹏、李海洋、孙楠、杨帆、杨涛、王衡、刘自方、边疆、王井玲、刘音、陈鹰鹏、陈邹琦、董庆南、王淑雅。

We also acknowledge the 12 participating professors and discipline leaders from Capital University of Physical Education and Sports. Their individual names are not listed in this release.

Scope, license and limitations

The full project contains 809 questions; this companion annotates only the existing 80-question public subset. Published answers and rubrics make this an exposed research/development resource, not an unseen held-out test. Disclose prior exposure and benchmark-specific optimization. The materials are not medical advice or evidence of deployment safety.

Questions and rubrics originate from the company repository's public release. Descriptions and scores are preserved; MCQ answer labels are normalized to lists. All content here is licensed under CC BY 4.0; credit MoveBench Team and cite the paper. The paper itself has a separate license.

中文说明

这是独立的答案与评分标准数据集，不重复发布题干。包含原公开 80 题的题号、30 道选择题答案、50 道开放题的 464 条 rubric。通过 problem_id 与原始题库关联；不含开放题参考答案。专家人数和工时为完整 809 题项目整体投入。