NewsReWrite: data, code, model checkpoints, and human annotations for LLM-guided headline rewriting without clickbait

Artifacts accompanying the paper LLM-guided headline rewriting for clickability enhancement without clickbait. The method steers Meta-Llama-3-8B-Instruct at inference time with the FUDGE paradigm, using two BERT guide models: a prefix clickbait scorer (negative guidance) and an engagement-attribute model (positive guidance), with a clickbait brake that decouples engagement from clickbait. This deposit contains: Generated data (newsrewrite_data.zip): the synthetic clickbait corpus, source-neutral headlines, and prefix train/val/test splits. Code and final results (newsrewrite_code_results.zip): dataset generation, guide-model training, FUDGE decoding, the DExperts/GeDi baselines, evaluation scripts, and the per-item result files behind every paper table and figure. Human annotations (newsrewrite_annotations.zip): the three-annotator study behind Section 4.10 — filled workbooks, per-rater CSVs, oracles, codebooks, and analysis outputs for rubric validation (C1), rewrite quality (C2), and pairwise engagement (C3). Guide-model checkpoints: the prefix clickbait guide and engagement-attribute guide (fine-tuned from bert-base-uncased), plus the independent DistilBERT clickbait detector (fine-tuned from distilbert-base-uncased) used for unbiased evaluation. Not included: third-party corpora (Chakraborty 2016, ISOT, Webis-17) are referenced with retrieval instructions rather than rehosted; the DExperts/GeDi baseline checkpoints derive from Llama-3.2-1B and are reproducible from the included scripts. See LICENSES.md and EXTERNAL_DATA.md.