Blocking Score and Commercialization Probability for European Patents - EPO CodeFest 2025

This dataset provides two novel patent value indicators computed for European Patent (EP) publications, developed as part of the EPO CodeFest 2025 on "Patent and IP Portfolio Valuation" by the Red Cube Coders team. Blocking Score The blocking score quantifies a patent's ability to limit the scope of subsequent patent applications. It is derived by: (i) measuring the scope of patent claims using information-theoretic surprisal computed with a causal language model (Qwen3-0.6B); (ii) calculating the scope degradation between application and granted claims; and (iii) fractionally attributing the degradation to focal patents cited as X or Y prior art during prosecution. The final score aggregates these attributions across all citing patents, with cross-language normalization (EN/DE/FR) and per-CPC-section sigmoid normalization. Commercialization Score The commercialization score estimates the probability that a patent protects a commercialized product or process. It is based on a positive-unlabeled (PU) learning model trained on the IPRoduct database (EPFL), which extracts patent-product correspondence from virtual patent marking (VPM) webpages. This is the first application of such a model to the full European patent population. Files included - blocking-score_blocking_scores.parquet — Per-patent blocking scores with temporal breakdowns (1,246,387 cited patents, 12 columns) - blocking-score_attribution_matrix.parquet — Per-citation-edge attribution detail (1,846,379 edges, 10 columns) - commercialization_final_dataset_raw.pkl — Per-patent commercialization dataset for EP granted patents, including patent metadata, CPC classifications, applicant/inventor information, claims, citation counts, and commercialization indicators (1,569,711 patents, 59 columns) - citations.csv — X and Y patent citations between EP patents from PATSTAT (2,450,384 rows, 18 columns) Source data - PATSTAT (EPO): Citations (X/Y categories), CPC classifications, publication metadata - EPAB API (EPO): Full-text patent claims (application and granted versions) - IPRoduct database (EPFL): Patent-product correspondence from virtual patent marking webpages - CPC classification titles: Cooperative Patent Classification hierarchy (Jan 2026 release) Models used - Qwen3-0.6B (causal LM): Per-token surprisal computation for claim scope measurement - PU learning classifier: Commercialization probability estimation from positive-unlabeled data Source code The full source code, including scripts, notebooks, and reproducibility documentation, is available at: https://github.com/rezaho/CodeFest-RedCubeCoders