watsonx.ai

 View Only

From Big to Smart: Cut AI Costs with smaller, customized Models Using InstructLab on watsonx.ai(private preview)

By Suhas Kashyap posted 10 hours ago

  

Today, we’re launching a private preview of model alignment with InstructLab in watsonx.ai. With this feature, customers can add their enterprise data, amplify that with synthetic data and alignment tune a small language model. This results in a smaller specialized model that is catered to an enterprise use case without sacrificing on the performance of the model. Earlier in the year, IBM and Red Hat announced InstructLab, a model alignment methodology to customize smaller models with the intent of reducing costs associated with operationalizing AI in enterprises. IBM watsonx.ai had also issued a statement of direction to work on bringing in this technology and adding enterprise workflows within the watsonx.ai platform.

While methods like prompt tuning, PEFT-based tuning, full fine-tuning, and RAG each play vital roles in adding enterprise data to AI models, they often come with limitations such as challenges with data preparation or risks of catastrophic forgetting. InstructLab on watsonx offers a breakthrough with its larger model alignment strategy, enabling enterprises to seamlessly and efficiently integrate their data in a structured way. By alignment tuning models to enterprise-specific needs, InstructLab ensures long-term knowledge retention without sacrificing accuracy or efficiency—unlocking powerful, cost-effective AI tailored for business success.

How does it work?

Built on IBM Research’s LAB methodology, and Red Hat Enterprise Linux AI,  InstructLab in watsonx.ai, offers end to end enterprise grade scalable workflows for model customization.

To get started, login to watsonx.ai using an IBM ID. Search for Secrets Manager from the IBM Cloud Catalog and create a new resource

Create a new Secrets Manager secret with the Github Personal Access Token (PAT)

Go to Tuning Studio. You’ll have 2 options: Prompt Tuning and InstructLab. Choose InstructLab to start customizing the model.

(Note: You’ll need a Standard plan and need to be allow-listed to access the functionality)

You can either import an existing taxonomy from Github or create your own taxonomy.

Taxonomy management

Choose the taxonomy you want to add your enterprise data to. Browse the nodes in the taxonomy and choose the node you want to add the enterprise data to.

Data ingestion

Give seed questions and answers. Optionally choose to ground the information. These markdown files must be converted from pdfs in an earlier step. Watsonx.ai makes use of the text extraction feature for the document conversion. This is an optional chargeable feature.

Synthetic data generation

Kick off synthetic data generation. This will kick off an agentic pipeline leveraging InstructLab’s teacher model which will generate synthetic data and a critic model which will filter out any data that has hate, abuse and profanity (HAP) in it. This will be limited to 3 successful runs in a month for an account.

You can review the synthetic data generated.

Alignment tuning

Kick off alignment tuning with InstructLab. This will queue up and trigger a job that will run on IBM Cloud backend. This will also be limited to 3 successful runs in a month for an account.

Model deployment and evaluation

Once the alignment tuning is complete, you will be provided an option to deploy the model within watsonx.ai. This will follow our existing custom foundation model deployment rules including pricing.

Aligned model can also be evaluated with the following metrics that will be provided at the end of the alignment phase: Loss function, MMLU, MT-Bench and PR-Bench. Optionally also run the provided python notebooks to run IBM Bluebench benchmarks against the aligned model.

How to get started?

Reach out to your IBM representative to get on the waitlist


#watsonx.ai
#TuningStudio
#PromptLab
#GenerativeAI

0 comments
6 views

Permalink