微调
在任务特定数据上继续训练预训练模型，以调整其行为。
概览
微调是在已经预训练好的模型上，用更小的、针对特定任务的数据集继续训练，使其权重朝着你期望的行为偏移。它解决了通用模型可能无法遵循内部风格、领域词汇或固定输出格式的问题。它之所以存在，是因为预训练成本高昂而适配成本不高：从预训练检查点出发，意味着适度的数据集就足够了。当目标是行为而非新事实时，它是检索的主要替代方案。
技术图谱
微调 如何与技术栈的其他部分相连。
使用与依赖
- Instruction Tuning technique Techniques Instruction Tuning based on 微调 High ✓ source
- Supervised Fine-Tuning technique Techniques Supervised Fine-Tuning based on 微调 High ✓ source
应用于
- Axolotl framework Frameworks Axolotl used for 微调 High ✓ source
- DeepSpeed framework Frameworks DeepSpeed used for 微调 High ✓ source
- Dropout technique Techniques Dropout used for 微调 High ✓ source
- Gradient Descent technique Algorithms Gradient Descent used for 微调 High ✓ source
- Learning Rate concept Algorithms Learning Rate used for 微调 High ✓ source
- LoRA technique Techniques LoRA used for 微调 High ✓ source
- MLflow tool MLOps MLflow used for 微调 High ✓ source
- PEFT library LLM Frameworks PEFT used for 微调 High ✓ source
- RLHF technique Techniques RLHF used for 微调 High ✓ source
- Regularization concept Algorithms Regularization used for 微调 High ✓ source
- Reinforcement Learning concept Algorithms Reinforcement Learning used for 微调 High ✓ source
- Synthetic Data technique Techniques Synthetic Data used for 微调 High ✓ source
- TRL library LLM Frameworks TRL used for 微调 High ✓ source
- Weights & Biases tool MLOps Weights & Biases used for 微调 High ✓ source
Classification
- Overfitting concept Algorithms Overfitting part of 微调 High ✓ source
替代方案
对比
与上述替代方案的逐项对比。