Dataset
PatentAP
A resource built for forecasting patent approval outcomes, introduced in research employing domain-specific fine-grained claim dependency graphs, with further documentation to be supplied later.
Data preview
Dataset card
PatentAP is a publicly available tabular dataset of U.S. patent application records released by the shangdatalab-ucsd group at UC San Diego. It is designed for the machine-learning task of predicting whether individual patent claims will be approved, and supports research on fine-grained claim-level classification.
What this dataset contains
The dataset aggregates roughly 1.95 million rows of patent claim records drawn from publicly filed U.S. patent applications, distributed across train, validation, and test splits. Each row corresponds to a single claim tied to a parent application, with text fields such as claim_input and abstract alongside numeric features like percentile and various citation counts. It was introduced alongside the paper 'Beyond Scaling: Predicting Patent Approval with Domain-specific Fine-grained Claim Dependency Graph.'
Structure and fields
Records are stored in Parquet and include identifiers such as applicationNumber, claim_idx, and filingDate, plus classification metadata like patentClassification and applicationTypeCategory. Outcome labels are encoded as the integer columns claim_label_101, claim_label_102, claim_label_103, and claim_label_112, which flag specific grounds of rejection or objection. Engineered graph and similarity features include component, is_closed, is_open, similarity_product, transitional_phrase, and citation statistics such as max_citations and mean_score.
How analysts use it
For alternative-data analysts, PatentAP can serve as a training corpus for NLP models that score the probability of patent allowance from claim text and examiner-style features, useful for diligence on IP-heavy issuers or venture-stage companies building patent portfolios. The presence of citation and similarity fields (max_score_x, max_score_y, max_article_citations) also makes it a candidate input for prior-art strength or innovation-quality screens. Because the data is openly hosted on the Hugging Face Hub under a permissive size category, it can be combined with other public patent or litigation feeds to build custom approval-risk signals.
Documentation & references 1
Related datasets
Browse moreSEC EDGAR 10-K & 10-Q financial statements for 2,600+ NASDAQ companies.
Real Estate listings (2.2M+) in the US broken by State and zip code
About 1.5 million US patent claims split into training and test partitions in CSV, organized for claim-level classification and summarization work (Apache 2.0). A large, ready-split corpus for IP-claims modeling.
ALEA Patent Claim Graph Reading questions about the dependency structure of real U.S. patent claim sets: which claim a dependent claim refers back to, the chain of claims it rests on, its independent claim, which and how many claims depend on a given claim, and yes / no questions about dependency. Every answer is computed deterministically from the claim text shown in the row. This dataset is part
A copyright-clean text collection released by the ALEA Institute as part of the KL3M Data Project, intended for training language models; full documentation is available via the linked paper and GitHub repository.
A component of the ALEA Institute's KL3M Data Project supplying training material cleared of copyright concerns. The dataset card is currently a placeholder, directing readers to the GitHub repo and project paper for full documentation.
Discussion
Questions or observations about this dataset? Join the conversation on X.
No comments yet. Start the conversation on X.