This chapter examines how fine-tuning converts the broad but shallowly specialized capabilities of general large language models into controllable performance for professional translation services. It frames translation and related tasks (lexical annotation, terminology handling, stylistic rendering) as structured, multi‑stage decision pipelines rather than unconstrained text generation. Parameter‑efficient techniques (e.g., Low-Rank Adaptation, quantization) enable cost‑effective adaptation of 7B–32B models while balancing privacy and deployment trade‑offs. Two signature case studies—full lexical annotation plus vernacular translation of Classical Chinese, and low‑resource Tangut translation via four‑line alignment and dictionary sense enumeration—demonstrate that chaining subtasks (annotation → literal rendering → idiomatic rendering) and constraining choices markedly reduce hallucinations and improve consistency even with small curated datasets. The chapter emphasizes deliberate data engineering: schema design (prompt/input/output separation), structured JSON formats, contrastive and staged samples, expert–model iterative loops, and governance for quality, compliance, and updates. Evaluation shifts from generic overlap metrics to task‑specific behavioral rubrics and structural format integrity. Fine‑tuning is justified when tasks decompose into bounded decision spaces; otherwise prompt design or few‑shot methods may suffice. Overall, structured fine‑tuning operationalizes expert knowledge, enhances controllability, and supports capability transfer to low‑resource scripts while acknowledging limits in highly creative scenarios.