Where Experts Disagree, Models Fail: model and datasets for detecting implicit legal citations in French court decisions

Trained retrieval model (a fine-tuned JuriBERT bi-encoder) and datasets (the 1,015-pair annotated benchmark and the chunked training corpus with TF-IDF features) accompanying the paper. Heavy companion to the anonymous code repository, which contains the full pipeline, the analysis notebooks, and a self-contained notebook that reproduces every table from the shipped predictions. Court decisions are sourced from Judilibre (French Ministry of Justice) under the Etalab Open Licence 2.0 and are pseudonymised at source.