This option lets you evaluate the model before training, which can help you judge the quality of the LLM backbone before fine-tuning.