CCF Database: A Machine-Learning-Annotated Corpus of 266,271 Canadian Climate Articles (1978–2024) — Apache Parquet mirror

The Canadian Climate Framing (CCF) Database is a comprehensive, machine-learning-annotated corpus of climate-change media coverage in Canada. It comprises 266,271 articles from 20 major Canadian newspapers (1978-2024) processed into 9,198,958 two-sentence analytical units (82.9% English, 17.1% French). Each unit is annotated across 65 hierarchical categories by 128 BERT and CamemBERT classifiers, with a macro F1 of 0.866 on a 1,000-sentence gold standard double-coded by an independent annotator (Gwet's AC1 = 0.894, Krippendorff's α = 0.698, Cohen's κ = 0.596 on the 400 blind sentences). Each category receives an A/B/C reliability tier summarising annotation quality from classifier performance and inter-coder agreement. The deposit ships six relational tables (bibliographic metadata, sentence-level annotations, named-entity rollups, article-level aggregates, per-category reliability tiers, and 9,462,845 BAAI/bge-m3 sentence-and-title embeddings). Raw newspaper text is excluded for copyright reasons; bibliographic coordinates (media, date, title, author, page_number) are sufficient for any researcher with institutional access to Factiva, Eureka.cc or ProQuest Canadian Major Dailies to recover the original sentences. This deposit accompanies a methodology paper currently under revision at Scientific Data (Nature Portfolio). This deposit is the Apache Parquet mirror of the canonical PostgreSQL edition (cross-referenced in Related identifiers). Each of the six relational tables is provided as a standalone .parquet file with ZSTD compression; the 1024-dimensional BAAI/bge-m3 embedding column is materialised as a list<float>, and JSONB entity arrays are serialised as UTF-8 JSON strings. The schemas are otherwise identical to the PostgreSQL edition. The Parquet bundle is readable natively by pandas, polars, R/arrow, DuckDB, and Spark without any database backend: import pandas as pd agg = pd.read_parquet('CCF_article_aggregates.parquet') emb = pd.read_parquet('CCF_sentence_embeddings.parquet') The HNSW index that ships with the PostgreSQL edition is not transferable to Parquet; brute-force cosine similarity remains tractable on the embedding column (≈ 9.46 M × 1024 float16). The full annotation pipeline, training data, manual-annotation JSONL, intercoder-reliability benchmark, methodology manuscript (LaTeX sources + PDF), and reproducibility scripts are bundled with this deposit as ccf_code_and_paper.tar.gz. The same materials are also available on the project's OSF companion deposit (10.17605/OSF.IO/Q5W47) and on the development mirror at GitHub.