I built code-repair training data and shipped the eval so you can rerun it (huggingface.co)

A new dataset called TrueSET has been introduced, which significantly enhances code-repair capabilities in AI models. A 7B parameter model, fine-tuned on this dataset, achieved a fix rate of 40% on hard-tier faults—compared to just 17% for base models—a remarkable increase that underscores the dataset's effectiveness in tackling complex, multi-file issues that traditional models struggle with. Each training example consists of proven issue reports, diagnosis, and verified fixes, making it a robust resource for evaluating code repair performance without relying on subjective assessments from LLM judges. The significance of TrueSET lies in its rigorous validation process, where every fault is checked against hidden verifiers to ensure only genuine fixes are recognized. The dataset contains 924 meticulously crafted examples, complete with executable proofs, allowing researchers to reproduce results at minimal cost. This transparent methodology paves the way for future advancements in AI-driven software debugging, emphasizing the importance of reproducible and verifiable training data in improving ML model performance on difficult coding tasks.