微调 — 人工智能

🎛️ 微调

在较小的特定任务数据集上进一步训练预训练模型。

预训练赋予模型通用知识，而微调则使其具备专门针对特定任务的能力。该过程会继续在一个规模较小、针对特定任务的数据集上进行训练，并调整权重，使模型在目标领域表现良好。

完全微调会更新每个参数，这对大型模型来说计算量非常大。而像 LoRA 和适配器这样的参数高效方法只更新一小部分参数，就能以极低的计算成本获得类似的结果。

常见的微调场景

微调也可能引发问题。模型可能过度拟合新数据，从而丧失通用能力。灾难性遗忘会抹去预训练中积累的知识。谨慎的学习率、数据混合和评估有助于降低风险。

Comments (3)

Hobbyist September 28, 2026

Made a big difference for my domain-specific chatbot. Base model was too generic on its own.
ML practitioner September 11, 2026

Taking a pretrained model and training it further on a smaller specific dataset. Much cheaper than training from scratch.
Researcher August 16, 2026

Learning rate and which layers to freeze are the usual knobs. Overfitting the small set is the main risk.