README.md

---
language:
- en
license: cc-by-sa-4.0
task_categories:
- image-to-text
- text-to-image
- image-classification
- zero-shot-image-classification
tags:
- patent
- design-patent
- technical-drawing
- wireframe
- multimodal
- retrieval
- reconstruction
- USPTO
dataset_info:
  features:
  - name: figure_id
    dtype: string
  - name: patent_id
    dtype: string
  - name: image
    dtype: image
  - name: patent_title
    dtype: string
  - name: caption
    dtype: string
  - name: drawing_description
    dtype: string
  - name: detailed_description
    dtype: string
  - name: brief_summary
    dtype: string
  - name: claims
    dtype: string
  - name: viewpoint
    dtype: string
  - name: figure_number
    dtype: int32
  - name: n_figures_in_patent
    dtype: int32
  - name: sibling_figure_ids
    sequence: string
  - name: reference_numerals
    sequence:
      struct:
      - name: numeral
        dtype: string
      - name: label
        dtype: string
  - name: locarno_class
    dtype: string
  - name: patent_date
    dtype: string
  - name: year
    dtype: int32
  splits:
  - name: train
    num_examples: 500
---

# Patent Wireframes

Structured patent figure dataset for multimodal understanding — generation, reconstruction, retrieval, and cross-modal association.

## Overview

Each record is a single patent figure paired with:
- Full patent text (drawing description, detailed description, brief summary, claims)
- IMPACT-generated figure caption
- Extracted reference numeral mappings (numeral → component label)
- Sibling figure links (other views of the same object)
- Classification metadata (Locarno codes, dates)

## Sources

- **Images + captions**: [IMPACT](https://huggingface.co/datasets/AI4Patents/IMPACT) (CC-BY-SA-4.0)
- **Patent text**: [PatentsView](https://patentsview.org/) (CC-BY-4.0)
- **Scope**: US design patents, initially 2022

## Schema

| Field | Type | Description |
|---|---|---|
| `figure_id` | string | Unique: `{patent_id}_{figure_number}` |
| `patent_id` | string | USPTO design patent number |
| `image` | image | Figure raster |
| `patent_title` | string | Patent title |
| `caption` | string | IMPACT-generated caption |
| `drawing_description` | string | Drawing description from patent text |
| `detailed_description` | string | Full detailed description |
| `brief_summary` | string | Brief summary |
| `claims` | string | Patent claims |
| `viewpoint` | string | front / side / perspective / etc. |
| `figure_number` | int32 | Figure index within patent |
| `n_figures_in_patent` | int32 | Total figures in patent |
| `sibling_figure_ids` | list[string] | Other figure IDs from same patent |
| `reference_numerals` | list[{numeral, label}] | Extracted component mappings |
| `locarno_class` | string | Locarno classification code |
| `patent_date` | string | Grant date |
| `year` | int32 | Grant year |

## Intended uses

- **Reconstruction**: predict figure from text description (or vice versa)
- **Retrieval**: text-to-figure, figure-to-figure, component-to-figure
- **Generation**: produce patent figures from structured descriptions
- **Multi-view reasoning**: predict unseen views given other views
- **Component understanding**: reference numeral grounding

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{patent_wireframes,
  title={Patent Wireframes: A Structured Dataset for Multimodal Patent Figure Understanding},
  year={2026},
  url={https://huggingface.co/datasets/midah/patent-wireframes},
}
```

Built on [IMPACT](https://openreview.net/forum?id=l0Ydsl10ci) (NeurIPS 2024) and [PatentsView](https://patentsview.org/).