QAQ: Bidirectional Semantic Coherence for Selecting High-Quality Synthetic Code Instructions

arXiv 2603.12165·medium signal

Addresses a core problem in synthetic code training data: low-probability outputs under IFD metrics conflate genuine task difficulty with model hallucination, making data quality filtering unreliable. QAQ adds a reverse coherence signal (how well the answer implies the question) to disambiguate, yielding cleaner synthetic instruction sets for code generation fine-tuning. Actionable for any team building code-specialized models on synthetic data pipelines.