Zero-shot g
The model can perform a new task without seeing any examples.
Zero-shot learning (zero-shot learning) is the ability to perform a new task based on a instruction given only in natural language without seeing any direct example of that specific type of task during training. For example, without any examples of a large language model "This customer's comment is positive, negative or neutrally classified", the model can produce a reasonable answer by using the general language and world knowledge that it learned during pre-training, even if it has never been trained in this specific format before. This ability was almost no in classic machine learning models; for each new task, it is necessary to have labeled training data.
Zero-shot learning is considered to be one of the capabilities of "emergent") that arises as the scale of large language models, and it is an important indicator that the model can not really generalize and do. Practically carries great value: expensive and time-consuming data labeling and model training processes for a new task, it becomes possible to get results directly with a instruction. However, zero-shot performance may not always give the highest accuracy; in complex or unusual tasks, the model shows several examples, the few-shot learning or model of fine-tuning, which is specially trained to the task often provides more reliable results.
Zero-shot learning is considered to be one of the capabilities of "emergent") that arises as the scale of large language models, and it is an important indicator that the model can not really generalize and do. Practically carries great value: expensive and time-consuming data labeling and model training processes for a new task, it becomes possible to get results directly with a instruction. However, zero-shot performance may not always give the highest accuracy; in complex or unusual tasks, the model shows several examples, the few-shot learning or model of fine-tuning, which is specially trained to the task often provides more reliable results.