InstructLab: A Novel Approach to Model Alignment

While the Granite models provide a strong foundation, IBM recognized that enterprises often have valuable proprietary data that they want to leverage. This is where the company's new InstructLab project comes into play.

InstructLab is an open-source tool that enables the collaborative addition of new knowledge and skills to language models, without the need for full retraining. Instead of relying solely on fine-tuning - which can overwrite a model's core capabilities - InstructLab uses a novel "alignment" technique to seamlessly integrate external data and instructions.