IBM have just announced their latest family of AI models, Granite 3.0, during their annual TechXchange event in Las Vegas.
The new models target enterprises looking to adopt generative AI technology, looking to capitalize on the rapid rise of workplace AI.
In this article we explore what the new Granite 3.0 models are, if they’re safe for enterprises to use and how to download them.
What is Granite 3.0?
Granite 3.0 is the latest family of large language models from IBM. LLMs are the powerhouses behind most of the latest AI innovations, from chatbots to content creators to code writers.
In a press release about Granite 3.0 IBM claims the new generation of models can outperform or match similarly sized models from leading model providers on many academic and industry benchmarks, showcasing strong performance, transparency and safety.’
The Granite 3.0 release introduces a family of new models built with different purposes in mind. For general use and language generation IBM recommends Granite 3.0 8B Instruct, Granite 3.0 2B Instruct, Granite 3.0 8B Base and Granite 3.0 2B Base. For a focus on guardrails and safety, IBM has created Granite Guardian 3.0 8B, Granite Guardian 3.0 2B whilst for a mixture-of-experts approach there is Granite 3.0 3B-A800M Instruct, Granite 3.0 1B-A400M Instruct, Granite 3.0 3B-A800M Base, Granite 3.0 1B-A400M Base.
The guardian models set out to provide a comprehensive set of risk and harm detection capabilities. This includes identified harm dimensions such as profanity, violence, jailbreaking and social bias. The models also provide a unique range of RAG-specific checks including context and answer relevance.
Read: What is a Large Language Model?
The Granite 3.0 models were trained on over 12 trillion tokens on data from 12 different languages and 116 different programming languages. The IBM developed the models by using a novel two-stage training method that used the results of thousands of experiments that were designed to optimize data quality, data selection, and training parameters.
Is Granite 3.0 Safe?
While most LLMs are trained on publicly accessible data, enterprise-level data often remains untapped due to privacy concerns.
IBM attempts to address these concerns by providing IP indemnity for enterprises that use WatsonX.ai, their AI driven data platform.
IBM envisions enterprises implementing an internal smaller Granite model in combination with their customisation tool, InstructLab, to take over task specific performance at a fraction of the cost of a larger model.
Alongside the release of the models, IBM have also created a technical report and responsible user guide that provide comprehensive guidance on the datasets the models have been trained on including the process behind filtering, cleansing and curation as well as demonstrations of each model's performance aligned with academic and enterprise benchmarks.
All of the new models are released under the permissive Apache 2.0 license. This open source software license allows developers to use, change and distribute the model for both commercial and non commercial purposes. This will provide flexibility and autonomy to the intended enterprise level consumers.
How To Download and Use Granite 3.0?
The whole range of IBM’s Granite 3.0 models are available to download HuggingFace, a online platform that hosts a library of pre-trained machine learning models, datasets, and tools.
1. Open your web browser and navigate to the Hugging Face website
2. Install the Hugging Face Transformers Library, which provides tools for working with LLMs
3. In the search bar at the top of the page, type "Granite 3.0" and press Enter. This will display a list of available Granite 3.0 models.
4. Choose the Granite 3.0 model you want to download.
5. Click on the model's name to go to its page.
6. Click 'use this model' and download to the appropriate library.
7. Select the format that is compatible with your development environment.
8. Save the files to your computer.
