For a long time, software developers have designed methodologies, processes, and tools that help them improve code quality and increase productivity. As an illustration, agile, test-driven development, code reviews, and CI/CD are actually staples within the software industry.
In “How Google Tests Software” (Addison-Wesley, 2012), Google reports that fixing a bug during system tests – the ultimate testing stage – is 1000x costlier than fixing it on the unit testing stage. This puts much pressure on developers – the primary link within the chain – to put in writing quality code from the get-go.
For all of the hype surrounding generative AI, code generation seems a promising solution to help developers deliver higher code fast. Indeed, early studies show that managed services like GitHub Copilot or Amazon CodeWhisperer help developers be more productive.
Nevertheless, these services depend on closed-source models that cannot be customized to your technical culture and processes. Hugging Face released SafeCoder a couple of weeks ago to repair this. SafeCoder is a code assistant solution built for the enterprise that provides you state-of-the-art models, transparency, customizability, IT flexibility, and privacy.
On this post, we’ll compare SafeCoder to closed-source services and highlight the advantages you may expect from our solution.
State-of-the-art models
SafeCoder is currently built on top of the StarCoder models, a family of open-source models designed and trained throughout the BigCode collaborative project.
StarCoder is a 15.5 billion parameter model trained for code generation in over 80 programming languages. It uses modern architectural concepts, like Multi-Query Attention (MQA), to enhance throughput and reduce latency, a way also present within the Falcon and adapted for LLaMa 2 models.
StarCoder has an 8192-token context window, helping it have in mind more of your code to generate recent code. It may possibly also do fill-in-the-middle, i.e., insert inside your code, as a substitute of just appending recent code at the top.
Lastly, like HuggingChat, SafeCoder will introduce recent state-of-the-art models over time, providing you with a seamless upgrade path.
Unfortunately, closed-source code assistant services don’t share information in regards to the underlying models, their capabilities, and their training data.
Transparency
According to the Chinchilla Scaling Law, SafeCoder is a compute-optimal model trained on 1 trillion (1,000 billion) code tokens. These tokens are extracted from The Stack, a 2.7 terabyte dataset built from permissively licensed open-source repositories.
All efforts are made to honor opt-out requests, and we built a tool that lets repository owners check if their code is an element of the dataset.
Within the spirit of transparency, our research paper discloses the model architecture, the training process, and detailed metrics.
Unfortunately, closed-source services stick with vague information, akin to “[the model was trained on] billions of lines of code.” To one of the best of our knowledge, no metrics can be found.
Customization
The StarCoder models have been specifically designed to be customizable, and we’ve got already built different versions:
- StarCoderBase: the unique model trained on 80+ languages from The Stack.
- StarCoder: StarCoderBase further trained on Python.
- StarCoder+: StarCoderBase further trained on English web data for coding conversations.
We also shared the fine-tuning code on GitHub.
Every company has its preferred languages and coding guidelines, i.e., easy methods to write inline documentation or unit tests, or do’s and don’ts on security and performance. With SafeCoder, we will enable you to train models that learn the peculiarities of your software engineering process. Our team will enable you to prepare high-quality datasets and fine-tune StarCoder in your infrastructure. Your data won’t ever be exposed to anyone.
Unfortunately, closed-source services can’t be customized.
IT flexibility
SafeCoder relies on Docker containers for fine-tuning and deployment. It is easy to run on-premise or within the cloud on any container management service.
As well as, SafeCoder includes our Optimum hardware acceleration libraries. Whether you’re employed with CPU, GPU, or AI accelerators, Optimum will kick in mechanically to enable you to save money and time on training and inference. Because you control the underlying hardware, you can too tune the cost-performance ratio of your infrastructure to your needs.
Unfortunately, closed-source services are only available as managed services.
Security and privacy
Security is all the time a top concern, all of the more when source code is involved. Mental property and privacy have to be protected in any respect costs.
Whether you run on-premise or within the cloud, SafeCoder is under your complete administrative control. You’ll be able to apply and monitor your security checks and maintain strong and consistent compliance across your IT platform.
SafeCoder doesn’t spy on any of your data. Your prompts and suggestions are yours and yours only. SafeCoder doesn’t call home and send telemetry data to Hugging Face or anyone else. Nobody but you must understand how and once you’re using SafeCoder. SafeCoder doesn’t even require an Web connection. You’ll be able to (and will) run it fully air-gapped.
Closed-source services depend on the safety of the underlying cloud. Whether this works or not on your compliance posture is your call. For enterprise users, prompts and suggestions will not be stored (they’re for individual users). Nevertheless, we regret to indicate that GitHub collects “user engagement data” with no possibility to opt-out. AWS does the identical by default but allows you to opt out.
Conclusion
We’re very excited in regards to the way forward for SafeCoder, and so are our customers. Nobody should should compromise on state-of-the-art code generation, transparency, customization, IT flexibility, security, and privacy. We imagine SafeCoder delivers all of them, and we’ll keep working hard to make it even higher.
In the event you’re curious about SafeCoder on your company, please contact us. Our team will contact you shortly to learn more about your use case and discuss requirements.
Thanks for reading!
