padjohn/cbert: C-BERT: Factorized Causal Relation Extraction (v3.0.0)

C-BERT v3.0.0 This release serves as the official software artifact for the paper "C-BERT: Factorized Causal Relation Extraction" (Johnson, 2026). It captures the exact state of the library used to produce the experimental results. 📄 Paper, 🤗 Models Reproducibility & Architectures This library implements both model architectures discussed in the paper within a unified framework, allowing for direct comparison: v3 (Factorized): The proposed architecture decomposing causal influence into Role, Polarity, and Salience heads. Usage: Default behavior (or explicitly via --architecture_version 3). v2 (Unified): The baseline architecture using a single 14-class classification head, re-implemented within this v3 framework. Usage: Reproducible via --architecture_version 2. ⚠️ Breaking Changes & Compatibility Span Recognition: This version utilizes a BIOES tagging scheme. Legacy Weights: Model weights trained on previous versions of this repository (using BIO tags) are not compatible with this release. Need to run old BIO weights? Please use the Legacy v2.0 Release. New Features Factorized Heads: Separates causal prediction into Role, Polarity, and Salience for 76.9% accuracy. LoRA Integration: Native support for Low-Rank Adaptation to fine-tune large base models. Indicator Taxonomy: Includes data/indicator_taxonomy.csv with 644 classified German causal indicators. Installation pip install . Citation @article{johnson2026cbert, title={C-BERT: Factorized Causal Relation Extraction}, author={Johnson, Patrick}, year={2026}, doi={10.26083/tuda-7797} }