Nvidia has teamed up with ServiceNow and Hugging Face to introduce a new family of open-access large language models (LLMs) for code generation.
The StarCoder2 platform was developed by the BigCode community keeping performance, transparency and cost-effectiveness in mind.
StarCoder2’s extensive reach is based on the training of 619 programming languages. The AI code generator comes in three flavors: 3 billion, 7 billion and 15 billion parameters.
StarCoder2 makes code generation possible for everyone
According to the announcement, the smaller variants are built to deliver strong performance while controlling computing costs. The smallest model is built together with ServiceNow and promises to match the performance of the original StarCoder option with 15 billion parameters, while the mid-spec option has the support of Hugging Face.
StarCoder2’s 15 billion parameter option is trained on Nvidia’s accelerated infrastructure.
The significant improvements mean that while greater performance can be unlocked by the Nvidia accelerated option, even the most basic variant is a significant advancement over previous generations and requires less advanced infrastructure.
Jonathan Cohen, vice president of applied research at Nvidia. “Nvidia’s collaboration with ServiceNow and Hugging Face introduces safe, responsibly developed models and supports broader access to responsible generative AI that we hope will benefit the global community.”
Additionally, StarCoder2 uses a new code dataset called The Stack v2, which includes new training techniques for understanding low-resource programming languages, mathematics, and program source code discussions.
In addition to performance and efficiency improvements, the organizations emphasize that StarCoder2 adheres to ethical AI practices, such as using responsibly sourced data licensed from Software Heritage’s digital commons. Developers can also opt out of having their data used for training.