IBM Corp. today introduced a new lineup of language models, the Granite series, that will become available as part of its watsonx product suite.

The Granite series is rolling out alongside several other new features. According to IBM, watsonx is receiving a tool that will make it easier for companies to create artificial intelligence training datasets. Another feature addition will make it easier to adapt neural networks for new tasks.

Introduced in May, watsonx is a suite of software products designed to help companies build generative AI models. It also promises to ease related tasks such as ensuring neural networks meet safety requirements.

The new Granite models that IBM debuted today will become available through a component of watsonx called watsonx.ai. According to the company, the latter offering provides tools that make it easier to build custom neural networks. Watsonx.ai also includes a collection of prepackaged AI models, which will be enhanced with the Granite series later this quarter. 

The Granite series includes two language models dubbed Granite.13b.instruct and Granite.13b.chat. IBM says they can summarize documents, perform “insight extraction” and generate text. The models were built using a 2.4-terabyte training dataset created by the company’s engineers. 

The two Granite models both have 13 billion parameters, which makes them compact enough to run on a single V100 graphics card from Nvidia Corp. The V100 is considerably less expensive than the chipmaker’s flagship H100 graphics card. As a result, the Granite series should theoretically be easier for companies to deploy than larger language models that require more sophisticated hardware to run.

“The initial Granite models are just the beginning: more are planned in other languages and further IBM-trained models are also in preparation,” Dinesh Nirmal, senior vice president of IBM Software, wrote in a blog post today. 

The Granite series is rolling out to watsonx.ai alongside two open-source AI models. The first is Llama-2, a general-purpose large language model from Meta Platforms Inc. IBM is also adding StarCoder, a neural network optimized for programming tasks that ServiceNow Inc. and Hugging Face Inc. released in May.

Besides a larger catalog of prepackaged neural networks, the latest version of watsonx.ai also includes new AI development features.

Creating a custom AI model requires a large amount of training data. In many cases, manually assembling that information can require a significant amount of time and effort. One way companies ease the workflow by automatically generating training data using software.

Such synthetic data, as it’s known, isn’t always as accurate as records created manually. But it’s often suitable for AI training.

According to IBM, watsonx.ai is receiving a built-in synthetic data generation tool. To use it, companies must upload a sample dataset such as a collection of purchase logs. Watson.ai can analyze those logs and generate synthetic records with similar features.

Adapting an already trained AI model to a new task usually requires retraining it, which can be a resource-intensive process. To address that challenge, IBM is equipping watsonx.ai with a parameter tuning tool. The tool makes it possible to optimize a neural network for new tasks without retraining it.

With parameter tuning, developers optimize an AI model by creating a second neural network that plays a kind of supporting role. The second neural network gives the AI model instructions on how to perform a given task. When those instructions are combined with users’ natural language prompts, the AI gains the ability to perform the task at hand more effectively than it otherwise could.

As part of the update it detailed today, IBM is also enhancing watsonx.data. That’s a component of the watsonx product suite built to help companies manage their AI training datasets.

According to IBM, the tool is being enhanced with the addition of a conversational interface. The interface will allow users to more easily visualize information stored in watsonx.data, refine it and find specific records. IBM is also adding a vector database optimized to hold embeddings, the mathematical structures AI models use to store their internal knowledge repositories. 

Image: IBM