SemChunk-C: Lightweight LLMs for Semantic C Code Segmentation

Semantic segmentation of C-family code has long been a challenge due to complex syntax, macro expansion, and irregular structures. Traditional chunking methods—fixed windows, heuristic splits, and syntax-based tools—fail to capture meaningful functional units, hurting retrieval and downstream LLM tasks. To solve this, a team of researchers (Boris Nazarov, Darya Frolova, Shaked Leibzirer, Pavel Kisilev) defined a set of code chunk categories and trained an LLM-based classifier that identifies chunk boundaries and assigns each chunk a descriptive functional attribute. Their approach leverages the LLM's semantic understanding to allow flexible, context-aware chunking.

The result is SemChunk-C, a family of lightweight models based on the first four Ettin encoders, with sizes ranging from 17M to 150M parameters. Despite their small footprint, these models robustly identify cohesive units—data structures, interface blocks, etc.—even in challenging real-world constructs like nested definitions and macros. In tests, SemChunk-C achieves high boundary accuracy and semantic coherence, matching or outperforming chunkers built on much larger code-oriented LLMs. The authors also validated improved downstream task performance on curated benchmarks. This work is published as arXiv:2606.23697 and is relevant for software engineering, AI, and programming languages communities.