Fine-tuning a model while simulating low-precision arithmetic to maintain accuracy after quantization.