cmuchancel/mechanical-utility-patents-100
Mechanical Utility Patents — 100 HTML Records A curated collection of 100 distinct granted U.S. utility patents from 100 patent families, saved as HTML for mechanical-engineering document analysis. Published by cmuchancel. Source disclosure: the patent text, bibliographic metadata, and drawings were obtained from Google Patents. Original PDFs and drawing assets are hosted on patentimages.storage.googleapis.com. These are existing published patent records, not inventions or… See the full description on the dataset page: https://huggingface.co/datasets/cmuchancel/mechanical-utility-patents-100.
Mechanical Utility Patents — 100 HTML Records
A curated collection of 100 distinct granted U.S. utility patents from 100 patent families, saved as HTML for mechanical-engineering document analysis. Published by cmuchancel.
Source disclosure: the patent text, bibliographic metadata, and drawings were obtained from [Google Patents](https://patents.google.com/). Original PDFs and drawing assets are hosted on patentimages.storage.googleapis.com. These are existing published patent records, not inventions or patent text authored by the compiler. ATTRIBUTION.md lists every source page and original PDF.
Companion app: Patent to SJS.
Contents and selection
The project selected apparatus and physical-mechanism patents covering gears, bearings, links, springs, cams, clutches, pumps, valves, grippers, presses, and related machine elements. Design patents and application-only publications were excluded. Selection did not require an active legal status. This is a small, purposive research collection, not a random or comprehensive sample.
The original archive records 2026-09-14 as its retrieval/build date. retrieved_utc preserves that timestamp; it is not a newly measured HTTP retrieval time. Release verification has its own timestamp.
Files and use
patents/: 100 readable HTML pages with the available abstract, description, claims, and embedded drawing images for offline reading.source-html/: 100 original saved Google Patents HTML pages. Online resources may require internet access.index.html: searchable offline catalog.catalog.json: metadata and provenance for all 100 records.data/patents.jsonl: 100 rows, each including complete reader HTML in thehtmlcolumn, for the dataset viewer and library.mechanical-utility-patents-100.zip: portable dataset copy. Extract it and openindex.htmlto browse.ATTRIBUTION.md,RIGHTS.md,SHA256SUMS, andresearch/: attribution, rights notes, checksums, and validation records.
from datasets import load_dataset
patents = load_dataset("cmuchancel/mechanical-utility-patents-100", split="corpus")
record = patents[0]
print(record["patent_id"], record["source_url"])
html = record["html"]There is one `corpus` split with 100 records. No train/validation/test partition or ground-truth SJS annotations are supplied. The companion app uses this as an input/reference collection; it has not trained a model on this release.
Record fields
Preparation and verification
The project preserved Google Patents pages, extracted metadata and patent sections, built readable HTML, and embedded downloaded drawings. Mechanism summaries and selection notes are editorial additions, visibly separated from reproduced patent sections. They are not official patent-office annotations or validated labels.
Release checks compare every saved source with its recorded SHA-256 digest, confirm 100 unique publications and families, compare reader abstract/description/claims text with the source after whitespace normalization, and decode embedded images. See release verification. Earlier archive validation is retained separately. These integrity checks do not establish model accuracy.
Intended uses and limitations
Intended for teaching, exploratory parsing, information extraction, and human-reviewed systems modeling. English, U.S. grants, selected mechanical topics, and the 25-per-category balance limit representativeness. Related designs or language can overlap even across distinct patent families.
Source OCR, formatting, equations, and editorial notes may contain errors. Compare original PDFs when exact content matters. Images are document content; the companion app's current text pipeline does not analyze drawing pixels. Public inventor and assignee names remain as bibliographic metadata. No additional private personal data was collected for this release.
This is not a legal-status, novelty, freedom-to-operate, or engineering-safety assessment. No task benchmark or model performance scores are claimed.
Rights and citation
No blanket dataset license is asserted. The compiler does not claim ownership of upstream patent text, drawings, or Google page material. See RIGHTS.md; source attribution is not a license grant.
When using this collection, cite the dataset and preserve per-patent source links:
cmuchancel. (2026). Mechanical Utility Patents — 100 HTML Records. Hugging Face. https://huggingface.co/datasets/cmuchancel/mechanical-utility-patents-100. Source records: Google Patents; archive date: September 14, 2026.
Documentation format
This card applies the provenance, intended-use, and limitations principles in the requested model-card guide, using dataset-card metadata because this repository contains a dataset. There are no model weights, training parameters, or evaluation results to report.
