Repository-level transpilation means porting an entire codebase from one language to another while keeping cross-file dependencies, imports, and shared logic intact. AlignCoder is a retrieval-augmented framework built for repository-level code completion, and its architecture maps cleanly onto that…

GPTQ is a widely used method for compressing large models to 4-bit weights while retaining most full-precision quality. GGUF is a container file format that packages weights at various quantization levels, not a compression algorithm like GPTQ or AWQ. Full-precision large models can have very large…

Three methods, three data situations. Supervised fine‑tuning (SFT) copies labeled examples. Direct preference optimization (DPO) aligns to ranked “better vs worse” pairs. Reinforcement fine‑tuning (RFT) runs a reward loop that scores outputs and pushes toward the good ones. Most reinforcement…