Hugging Face, a leading AI startup, today announced the launch of SafeCoder - a new AI-powered code assistant solution designed specifically for enterprise use. Importantly, SafeCoder is more than just a machine learning model; it's a complete end-to-end commercial offering by Hugging Face tailored for enterprises.
SafeCoder aims to boost software developer productivity through code auto-completion, while addressing key enterprise requirements around security, privacy, and compliance.
Built on Hugging Face's StarCoder family of open source code models, SafeCoder is designed to be fully self-hosted on a company's own infrastructure. During training and deployment, a customer's proprietary codebase never leaves their virtual private cloud or exposure to any third party, including Hugging Face.
Hugging Face has addressed a significant pain point for enterprises that want to benefit from AI code assistant technologies built on LLMs. Tools such as GitHub Copilot, have been shown to significantly improve developer productivity and happiness. However, using closed-source LLMs for internal code assistance exposes organizations to a raft of security and compliance issues, including the risk of exposing intellectual property externally.
SafeCoder sidesteps these challenges by empowering enterprises to fine-tune their own open-source LLMs on their private codebases. Hugging Face provides the framework and tooling for organizations to prepare their own training datasets, fine-tune models like StarCoder, and deploy them privately.
With global regulations around machine learning models and datasets still evolving, SafeCoder places a heavy emphasis on compliance. It builds on the legacy of BigScience, another open-source project, to implement novel techniques specific to the code domain, such as consent mechanisms and commercially permissible license filtering.
To implement SafeCoder, Hugging Face works closely with customers during both setup and deployment phases. The company provides containers, scripts, and examples to prepare a training dataset and tailor the LLMs to a customer's specific needs. The solution is optimized to leverage various hardware configurations, including NVIDIA Ampere GPUs, AMD Instinct GPUs, Habana Gaudi2, AWS Inferentia 2, Intel Xeon Sapphire Rapids CPUs and more. It currently integrates with popular IDEs including VSCode and IntelliJ.
SafeCoder is launching in collaboration with VMware, allowing their enterprise customers to deploy it on VMware Cloud infrastructure. "Our collaboration with Hugging Face around SafeCoder fully aligns to VMware’s goal of enabling customer choice of solutions while maintaining privacy and control of their business data," said Chris Wolf, Vice President of VMware AI Labs. VMware has published reference architecture with code samples to enable quick deployment.
With SafeCoder, Hugging Face provides enterprises with a managed solution and an easy way to tap into the productivity promise of AI code assistants, while maintaining control, security, and compliance. It aims to be an alternative to closed source solutions lacking transparency or customizability.