For its data and AI platform, Watsonx, which was released in July, IBM has announced new generative AI models Granite.13b.instruct and Granite.13b.chat. The Granite series models are large language models (LLM) that support insight extraction, content creation, and summarization.
The “Decoder” architecture is used by IBM’s Granite series multi-size foundation models, which were unveiled on September 7. These models apply generative AI to language and coding applications.
For the Granite series, which should be ready this month, IBM intends to provide a thorough list of the data sources as well as a description of the data processing and filtering methods carried out to obtain training data.
For code generation on Watsonx.ai on IBM Cloud, IBM is also providing models from third parties, such as Meta’s Llama 2-chat 70 billion parameter model and the StarCoder LLM (large language model).
The enterprise-focused data lake of IBM is utilized to train the Watsonx.ai models. In order to deploy models and applications for governance, risk assessment, compliance, and bias mitigation, the company claimed to have built a training process that includes rigorous data collecting and makes use of control points.
According to IBM, every dataset that is used for training goes through a specified governance, risk, and compliance (GRC) assessment procedure because the Granite models will be made accessible to clients for customization to fit their specific applications. Governance practices are in line with IBM AI Ethics guidelines for adding data to the IBM Data Pile, said the company.