Training Data Quality Issues Cause Code Generation Failures in LLMs

A comprehensive literature review encompassing 114 primary studies explores the transmission of training data quality problems into failures in code generation within large language models. This research introduces a cohesive taxonomy that classifies generated code quality concerns across nine dimensions, as well as training data quality issues into both code and non-code attributes. It articulates a causal framework that outlines 18 common mechanisms of propagation and compiles advanced detection and mitigation techniques. The findings identify the origins of logical errors and security flaws in training datasets, contesting the notion that these issues are exclusively tied to model-level deficiencies.