Explore IBM foundation models that were tuned by using InstructLab. Explore the foundation model taxonomy to find knowledge gaps that you can fill by contributing skills.
IBM watsonx.ai supports foundation models that use a novel alignment-tuning method from IBM Research. Large-scale Alignment for chatBots, or LAB, is a method for adding new knowledge and skills to existing foundation models by generating synthetic data for the new capability, and then using the generated data to instruction tune the foundation model.
Instead of building a new foundation model from scratch, or creating a derivative model with fine-tuning, you can use IntructLab to augment an existing foundation model with the capabilities needed for your use case.
The following foundation models were tuned by using the alignment tuning method from InstructLab:
- granite-20b-multilingual
For more information about these foundation models, see Supported foundation models.
Exploring knowledge and skills
When a foundation model is trained with contributed knowledge and skills, a model taxonomy page is available for the model. The taxonomy shows a hierarchy of the knowledge and skills that were contributed to the foundation model, where each end node represents a contributed skill.
From the model card for the foundation model in Prompt Lab, click Training taxonomy to explore the knowledge and skills that were added to the model by using the LAB tuning method.
- A skill teaches the model to do something, such as write a promotional email or summarize a report. When you contribute a skill, you provide at least five seed examples that are used to generate synthetic data for training.
- Knowledge is the data and facts that a model needs to more accurately answer questions. When you contribute knowledge, you provide grounding data and question-and-answer pairs related to the data.
If a skill that you need for your use case is not available from the foundation model, build and contribute the skill.
Learn more
- Red Hat blog: What is InstructLab?
- IBM Research: A faster, systematic way to train large language models for enterprise
- IBM Research: What is AI alignment?
- Research paper on LAB
Parent topic: Supported foundation models