Fine-tuning
Retraining a pre-trained model with special data for a specific task.
Fine adjustment (fine-tuning) is a pre-trained model on a large and general-purpose data set, a much smaller and specific task or space-specific data set. Basic idea is to greatly protect the general language information, conceptual relations and reasoning patterns that the model learned during pre-training, but slightly reassemble to fit the target task of weights. For example, a general-purpose language model can be made more responsive to legal terminology and document format by passing fine adjustment on legal documents.
The biggest advantage of fine setting is efficiency: train a model from scratch requires millions of dollars and huge data clusters, fine setting can provide much less data, calculation strength and noticeable performance gains over time. "parameter-efficient" fine adjustment techniques, such as LoRA (Low-Rank Adaptation), not all the billions of weight of the model, but only updating a small layer of additional reduces memory and cost needs. With fine adjustment, prompt engineering and RAG, a model is one of three basic methods to adapt to a specific use scenario; which depends on the nature of the task to be preferred, the amount of data and budget.
The biggest advantage of fine setting is efficiency: train a model from scratch requires millions of dollars and huge data clusters, fine setting can provide much less data, calculation strength and noticeable performance gains over time. "parameter-efficient" fine adjustment techniques, such as LoRA (Low-Rank Adaptation), not all the billions of weight of the model, but only updating a small layer of additional reduces memory and cost needs. With fine adjustment, prompt engineering and RAG, a model is one of three basic methods to adapt to a specific use scenario; which depends on the nature of the task to be preferred, the amount of data and budget.