Dataset: Matehmatical Software Communties (extracted from zbMATH Open data)

This dataset contains the raw bibliographic metadata, derived co-usage graph, and trained classification models produced by research on mathematical software ecosystems, accompanying the paper "Multi-Label Classification of Mathematical Software Communities Using Graph-Based Modeling of Structured Domain Metadata" (ICMS 2026). Compiled by FIZ Karlsruhe, it provides structured information on mathematical software packages, the research articles that cite them, the resulting software co-usage graph, and the classifiers trained to predict software-community membership from article metadata. The dataset covers approximately 32,000 software entries and over 628,000 citation links, offering a comprehensive view of the computational mathematics landscape as curated by zbMATH Open editors. What do I need? If you want to... Download Use the two headline classifiers (Neural Network, Logistic Regression) Top-level files only: msc_model.pth/title_model.pth, msc_model.joblib/title_model.joblib, mlb_input.pkl, mlb_output.pkl Use or reproduce all six classifiers from the paper's results table The above plus artifacts_cache.zip (SGD, Passive-Aggressive, Multinomial NB, and Complement NB checkpoints live in data_cache/ inside it) Reproduce the full pipeline from scratch (graph → communities → training) software_metadata.zip, articles_metadata_collection.tar.gz, software_graph.edgelist, processed_graph.edgelist, cluster_sizes.txt, grouped_data_by_paper_id.csv, grouped_data_by_paper_id_partitions.pkl, plus the companion code Inspect the raw per-entry JSON records software_metadata.zip (per-software records) or articles_metadata_collection.tar.gz (per-article records) Audit or resume a specific pipeline step (preprocessing, oversampling, the autoencoder) artifacts_cache.zip (large, ~3.8 GB) Just inspect the software co-usage graph or communities software_graph.edgelist, cluster_sizes.txt Dataset Contents: 1. Software Metadata (software_metadata.zip) Canonical references and descriptions for ~32,000 mathematical software packages. Internally organized under a temp_folder_software_metadata/ subfolder, with one JSON file per software entry (e.g. 47323.json), keyed by numeric swMATH software ID. Each record includes: id, name, authors, classification (MSC codes), description, homepage, keywords, dependencies, programming_languages, operating_systems, license_terms, source_code, orms_id (external ORMS registry identifier, where available), related_software (cross-references to other software IDs/names), standard_articles (short citing-article stubs), articles_count, and zbmath_url. Coverage spans disciplines from numerical analysis and statistics to topology and number theory. 2. Raw Article Metadata (articles_metadata_collection.tar.gz) The unprocessed bibliographic layer underlying the paper-level dataset in item 3 below: one JSON record per citing article, keyed by numeric article ID, under a data/ subfolder. Each record includes: title (with original/subtitle/addition variants), year, zbmath_url, contributors (author list), source (publication venue details — series, publisher, volume, ISSN — plus a formatted citation string), references (the article's own cited works, each carrying its own zbMATH document ID, MSC codes, and year), and states (status flags, e.g. "has references" / "is cited"). 3. Processed Article-Level Dataset (grouped_data_by_paper_id.csv, grouped_data_by_paper_id_partitions.pkl) Aggregated from the raw records in item 2 above: one row per paper, joining paper titles, authors, publication venues and years, expert-assigned MSC codes (63 top-level classes), and the specific software packages each paper cites. Each paper is additionally mapped to the software community partition(s) of the tools it cites. 4. Software Co-Usage Graph (software_graph.edgelist, processed_graph.edgelist, cluster_sizes.txt) Undirected graph over the ~32,000 software packages, with edges derived from co-citation in the article corpus. software_graph.edgelist: raw co-citation graph, pre-pruning. processed_graph.edgelist: graph after edge-weight pruning (τ = 2), used as input to community detection. cluster_sizes.txt: node counts per Louvain partition (197 communities, resolution = 1.0, seed = 42, modularity Q = 0.6852). 5. Trained Classification Models (msc_model.pth, msc_model.joblib, title_model.pth, title_model.joblib, mlb_input.pkl, mlb_output.pkl) Two parallel article representations — MSC codes and SPECTER title embeddings — each used to predict software-community membership. .pth files: trained PyTorch neural network classifiers (MSC-code and title-embedding branches). .joblib files: trained scikit-learn Logistic Regression classifiers (MSC-code and title-embedding branches). mlb_input.pkl / mlb_output.pkl: fitted MultiLabelBinarizer objects for the input feature space and the 196 non-empty community labels. 6. Cached Intermediate Artifacts (artifacts_cache.zip, 40 files, ~3.81 GB uncompressed) autoencoder_cache/: the SPECTER-embedding autoencoder (full encoder+decoder and encoder-only checkpoints). embeddings_cache2/: the original 768-dimensional SPECTER title embeddings prior to compression, chunked across 9 files (164,552 papers total). data_cache/: preprocessed features, train/test splits, oversampled training data, and cached checkpoints for the SGD, Passive-Aggressive, Multinomial NB, and Complement NB classifiers on both representations, plus a standalone MLP baseline. Not required if you only want the two headline models in item 5; needed for the other four classifiers or for reproducing/auditing the pipeline end-to-end. Companion code: https://github.com/swMATH/Paper2Community