MAGAR-Bench — Test inputs

This repository distributes only the public Test inputs of MAGAR-Bench: 1,000 samples, six assembly families, and 1,140 unique PNG files. The paper's full benchmark contains 4,000 samples (3,000 Train and 1,000 Test); the Train split is not included in this release.

Task family	Test samples
Assembly Decision	238
Atomic Action	662
Minimal Rework	100
Total	1,000

The six Test families are gearbox_structure, luban_12_pillars, parametric_box_lock_01, parametric_tower_lock_10, spherical_interlocking_structure, and triangle_dovetail_structure. They are disjoint from the original Train families.

Contents

All upload files are at the repository root; there is no need to drag thousands of image files into the web uploader.

File	Purpose
test.jsonl	1,000 Test inputs; no assistant messages or reference answers
images.zip	Exactly the 1,140 PNG files referenced by Test; paths start with images/
image_manifest.json	SHA-256 and byte size of each archived image
schema.json, task_contract.md	Input fields and model output contracts
release_manifest.json	Release version, counts, source hash and export changes
load_test.py	Standard-library loader with validated image extraction
verify_release.py	Record, archive, boundary and checksum validation
LICENSE, ASSET_LICENSE.md	License scope and image/Unity rights status
MANIFEST.sha256	Checksums of the shipped files

Load a downloaded repository

Python 3.11 or newer is required; no third-party packages are needed:

python verify_release.py
python load_test.py

The loader extracts the image archive into a temporary cache outside the repository and resolves each record's image paths. To choose a persistent cache location:

python load_test.py --extract-dir ../magar_test_image_cache

From Python, with this repository on the module search path:

from load_test import load_test

records = load_test()
sample = records[0]
messages = sample["messages"]  # public user prompt, in original order
image_paths = sample["images"]  # absolute paths to extracted PNG files

Alternatively, read the table using Hugging Face Datasets:

from datasets import load_dataset
ds = load_dataset("json", data_files={"test": "test.jsonl"})

This last example loads only the table. It does not download/extract the image archive. Image paths in test.jsonl are relative to the archive extraction directory. The Dataset Viewer is configured for the Test table; image path strings are not embedded Image features and do not provide inline image previews. Extract images.zip before using the PNG files locally.

Evaluation boundary

Reference answers, locked Test gold, training records, human-evaluation records, LLM-as-Judge outputs and experimental results are not included. verify_release.py checks release integrity; it is not a benchmark scorer. This package alone cannot compute answer-based benchmark scores; scoring requires the separately maintained private references and evaluator.

Use the user messages and images as model inputs. Metadata is for indexing and analysis. Predict one JSON object per sample according to task_contract.md.

Version and provenance

Release magar-bench-test-inputs-v1 is a separately versioned public export of the frozen hardened-v3 Test inputs. Sample IDs, row order, user messages, image order and image bytes are preserved. It is not a new benchmark split.

Metadata is reduced to the allowlist in schema.json. The export removes semantic high-level task IDs, internal provenance paths, reviewer fields and construction/audit flags. Removed-field counts are recorded in release_manifest.json. This is an explicit export change; the original frozen source files remain unchanged.

Some legacy Assembly Decision prompts mention high_level_task_id before a later explicit vocabulary clause excludes it. The three output keys in task_contract.md and that later clause are authoritative. We preserve the original user messages and document this inconsistency rather than silently rewriting the frozen prompts.

License and attribution

Dataset content and cleared images use CC BY-NC 4.0; original Python helpers use Apache-2.0. See LICENSE and ASSET_LICENSE.md. Asset ownership and third-party redistribution clearance are not established by an integrity check. The inherited maintainer asset attestation remains pending.

The associated manuscript is MAGAG Multimodal Agents for GUI-Based 3D Mechanical Assembly and Minimal Rework. A verified author list and public paper identifier were not available during this export; no arXiv identifier or author citation is asserted here.