Model customization with Instructlab

Democratizing LLM and SLM customization

InstructLab, a feature within the Tuning Studio of watsonx.ai, will enable AI developers, data scientists and SME’s of any domain to customize LLMs and SLMs to increase model performance and productivity. InstructLab will offer a full-scale alignment-tuning experience, equipped with data lineage and evaluation features, and the computational infrastructure needed to handle large-scale AI development. By bridging these two workflows—low-cost, accessible training with full-scale enterprise deployment—InstructLab can help businesses start small and scale up seamlessly as their AI needs grow.

With InstructLab, users gain a powerful AI customization tool that not only understands their unique use cases but will also evolve with them as their business and use of AI scale. By using InstructLab in watsonx.ai, businesses can create a robust, end-to-end AI solution that can take full advantage of enterprise grade model customization, AI Agents, foundational models, RAG capabilities and many more.

Why choose InstructLab for model customization?

Unified and collaborative experience

Technical and non-technical users can contribute to the model customization process, and then reuse those knowledge and skills across models

Rapid iteration

Customize models with enterprise data in a matter of hours not months

Cost savings

Infer a small specialized model, not a larger generic model

Model customization tools

watsonx.ai will offer advanced enterprise capabilities and workflows for model customization with InstructLab. Read the statement of direction here.

Data ingestion pipeline

Ingest thousands of documents and manage complex ingestion pipelines (e.g., masking, chunking, and filters). Process multiple document formats (e.g., PDF, PPT, TXT, DOC) from various data sources.

Taxonomy management

Continuously build a taxonomy of enterprise knowledge and skills visualized in an easy to navigate tree structure. Watsonx.ai will provide UI, CLI, API and SDK for taxonomy creation.

Synthetic data generation

Amplify the taxonomy with InstructLab agentic synthetic data generation. watsonx.ai will provide UI, CLI, API and SDK for generating synthetic data.

Alignment tuning

Align the model with generated synthetic data in a multi-phased alignment technique. watsonx.ai will provide UI, CLI, API and SDK for alignment tuning.

Data lineage

Lineage tracking of the taxonomy, synthetic data and the aligned model

Model evaluation

Evaluate both pre-aligned and aligned models with Python notebooks to run IBM Bluebench benchmarks. Standard open benchmarks such as Loss function, MMLU, MT-Bench and PR-Bench.

Model choice in InstructLab

Leverage trusted, performant and cost-effective models optimized for your use cases

Choose your model

Power AI applications using our library of third-party and IBM-built Granite models suitable for AI workflows or bring your own custom foundation model to the platform.

Learn more about Granite

Explore foundation models in watsonx.ai

Take the next step

Ready to transform your business with InstructLab?

Start your free trial today (50,000 free tokens)

Explore the watsonx developer hub

Footnotes

IBM’s statements regarding its plans, directions and intent are subject to change or withdrawal without notice at IBM’s sole discretion.