Replication package: Small Language Models for Triaging Breaking Dependency Upgrades — An Empirical Study on 571 Reproduced Upgrade Failures
Code, data and predictions for the paper. Includes the BUMP-derived dataset builder, log condensation, rule and learned baselines, retrieval, prompting and QLoRA fine-tuning code for five small language models and a 32B reference, evaluation and statistics scripts, a Google Colab notebook for the GPU experiments, and e...