fine-tuning-workflowslisted

# Fine-Tuning Workflows > Fine-tuning is the heaviest tool in the LLM toolbox — costly to > do, costlier to maintain, and irreversible in the sense that the > base model evolves under you. Use it when prompting + RAG have > demonstrably plateaued, not before. Then do it properly. ## Purpose Fine-tuning adjusts a model's weights on a curated dataset to make it better at a specific behaviour: a domain vocabulary, an output schema, a writing style, a structured reasoning pattern, a classification task at lower cost than a frontier model. Done well, a fine-tuned smaller model can match or beat a much larger prompted model on the target task at a fraction of the inference cost. Done poorly — sloppy data, wrong objective, no eval — it produces a model that LOOKS better on the training distribution and fails silently on real traffic. This skill names the decision framework, the dataset discipline, the training-time choices, the evaluation gates, the safety re-tuning steps, the deployment pattern, and the maintenance overhead that every fine-tuning project incurs. NOT in scope: pretraining new foundation models from scratch (rare; specialised teams only); zero-shot prompt engineering (see `prompt-engineering`); retrieval-augmentation (see `rag-design`); classical ML model training (see `ml-model-selection`). ## Standards Cited - **Ouyang L., et al. (2022)** — "Training language models to follow instructions with human feedback" (InstructGPT — the SFT - RM + PPO recipe), Ne