Adapt pretrained models with supervised fine-tuning, adapters, low-rank methods, quantization-aware techniques, and carefully constructed training data.
