midah/patent-wireframes

TechnicalWireframeGym

Benchmark dataset for evaluating VLMs on US design patent figure conventions.

GitHub: hamidahoderinwale/TechnicalWireframeGym

Built from IMPACT (3.6M design patent figures, 2007–2022) enriched with PatentsView drawing descriptions.

Contents

Path	Description
data/enriched_{year}.parquet	16 years of figures with drawing descriptions (2007–2022)
data/eval_sample.parquet	800-patent evaluation split
embeddings/{year}_vitl14.parquet	CLIP ViT-L/14 embeddings, all 16 years (~2.5M figures)
results/*.json	Eval results per task and condition
status/embeddings.json	Provenance: completion records per year

Enriched parquet schema

Field	Type	Description
patent_id	string	USPTO design patent number
patent_title	string	Patent title
figure_number	int	Figure index (0-based)
n_figures_in_patent	int	Total figures in patent
drawing_description	string	Drawing description from PatentsView
locarno_class	string	Locarno classification
patent_date	string	Grant date
year	int	Grant year

Benchmark tasks

Task	Input	Metric	Gemma 4 31B
T1 Viewpoint ID	Vision / text FIM / vision+text	Accuracy	A:23% C:65% D:22%
T2 Figure retrieval	CLIP embeddings	R@1	34.2%
T3 Intent elicitation	Text	MC acc / F1	100% / 0.26
T4 Agent-agent	Structured schema	Accuracy	44%

Citation

@dataset{technicalwireframegym2026,
  title={TechnicalWireframeGym: Evaluating VLMs on Technical Drawing Conventions},
  author={Oderinwale, Hamidah},
  year={2026},
  url={https://huggingface.co/datasets/midah/patent-wireframes}
}