Instruction Tuning
Fine-tuning a model on examples of instructions paired with ideal responses.
Instruction tuning (also called supervised fine-tuning, SFT) trains a pre-trained base model on human-written instruction-response pairs. This transforms the raw next-token predictor into an assistant that follows user instructions. All commercial chat models — ChatGPT, Claude, Gemini — have been instruction-tuned on top of their pre-trained base.
Términos Relacionados
The initial large-scale training phase where a model learns language from massive text corpora.
Continuing to train a pre-trained model on domain-specific data to adapt its behavior.
Reinforcement Learning from Human Feedback — training models to align with human preferences.
Anthropic's method of training models to self-critique and revise outputs using a set of principles.