MAGAR-Bench

MAGAR-Bench is a frozen multimodal benchmark for GUI-based 3D mechanical assembly and minimal rework. The final paper defines three mutually exclusive task families: Assembly Decision, Atomic Action, and Minimal Rework.

This repository contains the public inputs and training records from the Experiment 2 freeze (frozen_experiment2, 2026-09-19): 4,000 records from 26 assembly families. Train and Test are assembly-family disjoint.

Split	Families	Assembly Decision	Atomic Action	Minimal Rework	Total
Train	20	736	1,764	500	3,000
Test inputs	6	238	662	100	1,000
Total	26	974	2,426	600	4,000

What is included

• data/train.jsonl: training records with user and assistant messages, metadata, and repository-relative image paths.
• data/test_inputs.jsonl: public Test inputs. Assistant answers and Test targets are removed.
• train.parquet and test.parquet: optional Pager-style flat tables generated from the canonical JSONL files with scripts/build_pager_style_parquet.py.
• data_types.py: column contract for the Parquet export.
• images/: the PNG files referenced by both JSONL files.
• contract_manifest.json: frozen split counts, ontology, contract version, and source hashes.
• schemas/schema.json: field and task-contract summary.
• docs/dataset_freeze.md: the freeze rules and data boundary.
• docs/task_contract.md: the output fields and allowed values for all tasks.
• scripts/load_dataset.py: a small local loader.
• scripts/verify_release.py: a stdlib-only integrity and leakage check.
• LICENSE: separate terms for dataset contents and original helper code.
• ASSET_LICENSE.md: image/Unity asset redistribution boundary and release attestation checklist.
• MANIFEST.sha256: SHA-256 hashes for every public release file.

The Atomic Action labels use five native GUI primitives: click, drag, hotkey, keypress, and type_text. Minimal Rework uses the validated recovery candidate set for each state; it does not claim unrestricted global optimality over all possible GUI action sequences.

Loading

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files={
        "train": "data/train.jsonl",
        "test": "data/test_inputs.jsonl",
    },
)

Each record contains an images list. Resolve those paths relative to the dataset repository root when constructing multimodal model inputs. The helper script provides the same behavior for a local checkout:

python scripts/load_dataset.py
python scripts/verify_release.py
python scripts/build_pager_style_parquet.py

The Parquet exporter is a representation conversion, not a new data freeze. Run it with --write only in an environment that has pyarrow installed. It keeps target_text for Train and leaves it null for Test.

Test labels and evaluation

The locked Test labels are intentionally outside this public directory. Keep the private evaluation artifact out of the public dataset repository. A public submission format and a gated evaluator should be distributed separately if leaderboard evaluation is enabled.

Release status

This directory is the canonical public dataset payload. LICENSE and ASSET_LICENSE.md define the proposed publication terms and the image/Unity asset boundary. Before publication, the maintainers must complete the asset rights attestation, add the final paper citation and repository URL, and publish any paper-supplement evaluation artifacts separately. Human evaluation and LLM-as-Judge records are intentionally outside this dataset payload.