FORGE

Dataset Description

FORGE is an engineering drawing visual question answering dataset of single-sheet CAD assembly drawings and compositional questions. Tasks connect feature annotations, geometry dimensions, balloons, bill-of-materials (BOM) rows, component names and quantities. It supports visual reasoning and evidence-centric research, including systems that acquire visual evidence, build explicit relations and execute compositional queries.

FORGE is a modified/derived dataset based on the Autodesk Fusion 360 Gallery Dataset. It is not the original Fusion 360 Gallery Dataset. Use is restricted to non-commercial research under the upstream Autodesk license. See License and Provenance before use.

Dataset Statistics

Split	Drawings	QA
Train	1,408	25,987
Validation	158	2,589
Test	158	2,731
Total	1,724	31,307

The frozen TEST taxonomy is Standard 2,464 and Hard 267: C1 89, C2 89, C3 58 and C4 31. Drawing IDs and exact image hashes are disjoint across splits; decoded RGB pixel hashes were also checked. Splits and all questions are preserved without performance-based curation.

Reasoning Taxonomy

Family	Reasoning
Standard	Balloon → BOM component, BOM quantity, feature → owner → BOM component/quantity, feature diameter, feature source view, and cross-feature diameter comparison.
C1	compare → component
C2	compare → quantity
C3	dual traversal → quantity comparison
C4	filter → quantity → diameter

C1–C4 form the Hard group. reasoning_family retains the finer native template names; metadata/taxonomy.json lists their mapping and counts. hop_count is the frozen symbolic annotation count, not measured model calls or a human annotation of reasoning difficulty.

Dataset Structure

The primary DatasetDict has train, validation and test splits with one row per drawing. Original PNG bytes are embedded once in Parquet per drawing; questions are a nested list, not duplicated image rows. Drawings sort by frozen drawing ID; QA order within a drawing preserves the native export order.

Drawing field	Type	Meaning
drawing_id	string	Frozen candidate_id, also the original image stem.
split	string	train, validation or test; native val is named validation.
image	datasets.Image	Original, unmodified drawing PNG with automatic image decoding.
drawing_width	int32	Native image width in pixels.
drawing_height	int32	Native image height in pixels.
qas	list of structs	All public QA attached to this drawing.

Nested QA field	Type	Meaning
qa_id	string	Frozen evaluation identity, forge:<native split>:<zero-padded source index>:<12-character request hash>.
global_qa_id	string	Original annotation identity: drawing ID plus fixed template/question identifier; contains no answer values.
question	string	Exact frozen question, excluding the export's <image> transport marker.
answer	string or null	Exact original display_answer; null on TEST.
canonical_answer	string or null	Exact original normalized reference value; null on TEST.
answer_text	string or null	Exact original natural-language answer text; null on TEST.
difficulty	string	Standard or Hard, determined only by frozen template family.
family	string	Standard, C1, C2, C3 or C4.
reasoning_family	string	Original fine-grained template family.
hop_count	int32	Original symbolic relation/hop count.
quality_grade	string	Original annotation quality grade; not independent human evidence verification.

manifests/{train,validation,test}_qa.jsonl provides one QA per line, with the same fields plus drawing_id and split, and no image bytes or image paths. The TEST JSONL omits all three answer fields entirely. Original reasoning explanations, symbolic traces and visible/source facts are not included in the public package.

Other public files: metadata/dataset_statistics.json contains counts; metadata/schema.json contains the feature schema and field policy; metadata/taxonomy.json contains taxonomy mappings; metadata/release_manifest.json inventories payload hashes; examples/loading.py demonstrates loading. LICENSE.md states the Modified Set terms and LICENSE_UPSTREAM.md preserves the Autodesk license.

Test Split Policy

The public test split contains drawings and questions but withholds reference answers and privileged source facts to preserve benchmark integrity. Answer columns in nested TEST QA are null; TEST manifest lines omit them. No reference reasoning, symbolic evidence trace or internal source facts are published. Labels remain unchanged in the private canonical source; they have only been withheld from this public projection. No public scoring server or label-request mechanism is provided by this release. Training and validation retain their original answer values.

Loading

from datasets import load_dataset

ds = load_dataset("herbwood27/FORGE")
drawing = ds["train"][0]
image = drawing["image"]  # PIL image, original resolution
qa = drawing["qas"][0]
print(drawing["drawing_id"], image.size, qa["question"], qa["answer"])
assert all(q["answer"] is None for q in ds["test"][0]["qas"])

# Flatten QA metadata without copying image bytes.
def iter_qa(split):
    for row in ds[split].remove_columns("image"):
        for qa in row["qas"]:
            yield {"drawing_id": row["drawing_id"], "split": split, **qa}

first_flat_qa = next(iter_qa("train"))

Use Image(decode=False) when inspecting embedded image bytes. Questions reference their drawing through drawing_id; join to the drawing table when batching QA.

Intended Uses

• Engineering drawing VQA for non-commercial research.
• Multimodal and compositional reasoning.
• Agentic and evidence-grounded reasoning research, including visual evidence acquisition and executable relation chains.

Limitations

The distribution consists of generated/rendered CAD assembly drawings. It is limited to single-sheet inputs and domain-specific mechanical drawing conventions, and is not representative of all industrial CAD or engineering drawings. Template-generated questions and symbolic annotation grades do not establish human-verified visual grounding. This release provides no human bounding-box ground truth. TEST answers are withheld. No model predictions, model checkpoints, training adapters, experiment traces or evidence crops are part of this dataset release.

License and Provenance

License title: Fusion 360 Gallery Dataset License (Modified Set). The Hub metadata uses the identifier fusion-360-gallery-dataset-modified-set.

FORGE is a Modified Set, derived from the Autodesk Fusion 360 Gallery Dataset, provided by Autodesk, Inc. It is not the original Autodesk dataset. The upstream source is Fusion 360 Gallery Dataset; its license is preserved in LICENSE_UPSTREAM.md. Publisher-declared upstream lineage and the frozen local export were audited; the export does not include a complete raw-CAD-to-drawing build ledger, so independent end-to-end reconstruction of that derivation is not claimed.

Access, use, reproduction and modification are restricted to non-commercial research. Downstream redistribution or modification must preserve restrictions at least as protective as the upstream license, including its non-commercial restriction, Autodesk warranty disclaimer and indemnification protections (items 7 and 8). Retain Autodesk / Fusion 360 Gallery Dataset attribution and associated notices, and impose these requirements on further recipients. The original upstream Dataset must not be redistributed in its entirety. See LICENSE.md and the complete upstream terms. FORGE is not offered under CC BY, MIT, Apache-2.0 or another permissive license, and no Autodesk endorsement is implied.

Citation

If you use FORGE in your research, please cite:

@misc{kwon2026forge,
  title        = {FORGE: Ontology-Grounded Compositional Reasoning over Mechanical Engineering Drawings},
  author       = {Kwon, Junehyoung and Kim, Mihyeon and Lee, Eunju and Lee, Dongjin and Lee, Seunghoon and Kim, YoungBin},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/herbwood27/FORGE}
}