IBM Launches "Granite": AI Foundation Models for Business
Zach Anderson Sep 07, 2023 16:34
IBM has introduced "Granite" foundation models for its watsonx.ai platform, utilizing generative AI for linguistic and coding applications. These models, specifically Granite.13b.instruct and Granite.13b.chat, have 13 billion parameters and are trained on diverse datasets. IBM emphasizes trust, transparency, and customization, and plans to introduce more models in different languages.
IBM has unveiled its latest advancement in the realm of artificial intelligence (AI) for business: a series of foundation models named "Granite." These models, designed for the watsonx.ai platform, aim to harness the power of generative AI for both linguistic and coding applications.
Granite Models: A Deep Dive
Developed by IBM Research, the Granite models, specifically Granite.13b.instruct and Granite.13b.chat, are built on a "Decoder" architecture. This underpins the capability of contemporary large language models to predict subsequent words in a sequence. With 13 billion parameters, these models are optimized for efficiency, fitting within a single V100-32GB GPU. Their design ensures reduced environmental impact while excelling in specialized business tasks such as summarization, question-answering, and classification.
The models are trained on a diverse range of datasets, amounting to 7 TB pre-processing and 2.4 TB post-processing, resulting in 1 trillion tokens. These datasets span various domains, including the internet, academia, coding, legal, and finance, ensuring the models are well-versed in industry-specific language and terminology.
Trust and Transparency: IBM's Commitment
IBM emphasizes trust and transparency in its AI development. The company's watsonx AI and data platform offers a comprehensive process for creating and testing foundation models and generative AI. This process encompasses everything from data collection to the deployment of models, with a keen focus on governance, risk assessment, bias mitigation, and compliance.
To ensure the integrity of the data used for training, IBM has implemented a rigorous governance, risk, and compliance (GRC) review process. Additionally, the company employs its "HAP detector," a language model developed to identify and eliminate hateful and profane content. This is part of IBM's broader strategy to mitigate the risks associated with generative AI.
Empowerment Through AI
IBM's vision for AI in business revolves around empowerment. The company believes that organizations should have the autonomy to customize their models in alignment with their values, using the tools provided by the watsonx platform. Furthermore, IBM ensures that businesses retain ownership and control over their data and models.
The Road Ahead
While the Granite models mark a significant milestone, they are just the beginning of IBM's journey in this space. The company has plans to introduce more models in various languages and is also working on other IBM-trained models. IBM has also announced its collaboration with Meta, offering early access to Meta's Llama 2-chat 70 billion parameter model. Additionally, IBM will host StarCoder, a comprehensive language model for coding.
In the coming weeks, IBM will also launch new features in the watsonx.ai studio, including the Tuning Studio and the Synthetic Data Generator. These tools will further enhance the capabilities of the platform, enabling businesses to harness the full potential of AI.
In conclusion, IBM's introduction of the Granite foundation models and the subsequent tools and features signifies a new era in AI for business, with endless possibilities on the horizon.
Disclaimer: The content of this article is for informational purposes only and is the exclusive property of Blockchain.News. While we aim for accuracy, we make no guarantees regarding its reliability or completeness.
Image source: Shutterstock