ServiceNow, Hugging Face, and NVIDIA Release Open Access LLMs

<p><b>SANTA CLARA<&sol;b>— <a href&equals;"https&colon;&sol;&sol;www&period;servicenow&period;com&sol;">ServiceNow<&sol;a>&comma; <a href&equals;"http&colon;&sol;&sol;www&period;huggingface&period;co&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">Hugging Face<&sol;a>&comma; and <a href&equals;"https&colon;&sol;&sol;www&period;nvidia&period;com&sol;en-us&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">NVIDIA<&sol;a>&comma; have announced the release of <a href&equals;"https&colon;&sol;&sol;huggingface&period;co&sol;bigcode&sol;starcoder2-15b&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">StarCoder2<&sol;a>&comma; a family of open‑access large language models &lpar;LLMs&rpar; for code generation that sets new standards for performance&comma; transparency&comma; and cost‑effectiveness&period;<&sol;p>&NewLine;<p>StarCoder2 was developed by the <a href&equals;"https&colon;&sol;&sol;www&period;bigcode-project&period;org&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">BigCode<&sol;a> community&comma; stewarded by<a href&equals;"https&colon;&sol;&sol;www&period;servicenow&period;com&sol;"> ServiceNow<&sol;a>&comma; the leading digital workflow company&comma; and <a href&equals;"https&colon;&sol;&sol;huggingface&period;co&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">Hugging Face<&sol;a>&comma; the most‑used open‑source platform where the machine learning community collaborates on models&comma; datasets and applications&period;<&sol;p>&NewLine;<p>Trained on 619 programming languages&comma; StarCoder2 can be further trained and embedded in enterprise applications to perform specialized tasks such as application source code generation&comma; workflow generation&comma; text summarization&comma; and more&period; Developers can use its code completion&comma; advanced code summarization&comma; code snippets retrieval&comma; and other capabilities to accelerate innovation and improve productivity&period;<&sol;p>&NewLine;<p>StarCoder2 offers three model sizes&colon; a 3 billion‑parameter model trained by ServiceNow&comma; a 7 billion‑parameter model trained by Hugging Face&comma; and a 15 billion‑parameter model built by NVIDIA with NVIDIA NeMo and trained on NVIDIA accelerated infrastructure&period; The smaller variants provide powerful performance while saving on compute costs&comma; as fewer parameters require less computing during inference&period; In fact&comma; the new StarCoder2 3 billion‑parameter model also matches the performance of the original StarCoder 15 billion‑parameter model&period;<&sol;p>&NewLine;<p>&OpenCurlyDoubleQuote;StarCoder2 stands as a testament to the combined power of open scientific collaboration and responsible AI practices with an ethical data supply chain&comma;” emphasized Harm de Vries&comma; lead of ServiceNow&&num;8217&semi;s StarCoder2 development team&comma; and co‑lead of BigCode&period; &&num;8220&semi;The state‑of‑the‑art open‑access model improves on prior generative AI performance to increase developer productivity and provides developers equal access to the benefits of code generation AI&comma; which in turn enables organizations of any size to more easily meet their full business potential&period;”<&sol;p>&NewLine;<p>&&num;8220&semi;The joint efforts led by Hugging Face&comma; ServiceNow and NVIDIA enable the release of powerful base models that empower the community to build a wide range of applications more efficiently with full data and training transparency&comma;&&num;8221&semi; said Leandro von Werra&comma; machine learning engineer at Hugging Face and co‑lead of BigCode&period; &OpenCurlyDoubleQuote;StarCoder2 is a testament to the potential of open‑source and open science as we work toward democratizing responsible AI&period;&&num;8221&semi;<&sol;p>&NewLine;<p>&&num;8220&semi;Since every software ecosystem has a proprietary programming language&comma; code LLMs can drive breakthroughs in efficiency and innovation in every industry&comma;” said Jonathan Cohen&comma; vice president of applied research at NVIDIA&period; &OpenCurlyDoubleQuote;NVIDIA’s collaboration with ServiceNow and Hugging Face introduces secure&comma; responsibly developed models&comma; and supports broader access to accountable generative AI that we hope will benefit the global community&period;”<&sol;p>&NewLine;<p><b>Fine‑Tuning Advances Capabilities with Business‑Specific Data<&sol;b><&sol;p>&NewLine;<p>StarCoder2 models share a state‑of‑the‑art architecture and carefully curated data sources from BigCode that prioritize transparency and <a href&equals;"https&colon;&sol;&sol;arxiv&period;org&sol;abs&sol;2312&period;03872" target&equals;"&lowbar;blank" rel&equals;"noopener">open governance<&sol;a> to enable responsible innovation at scale&period;<&sol;p>&NewLine;<p>The foundation of StarCoder2 is a new code dataset called <a href&equals;"https&colon;&sol;&sol;huggingface&period;co&sol;datasets&sol;bigcode&sol;the-stack-v2&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">The Stack v2<&sol;a> which is more than 7x larger than <a href&equals;"https&colon;&sol;&sol;huggingface&period;co&sol;datasets&sol;bigcode&sol;the-stack" target&equals;"&lowbar;blank" rel&equals;"noopener">The Stack v1<&sol;a>&period; In addition to the advanced data set&comma; new training techniques help the model understand low‑resource programming languages &lpar;such as COBOL&rpar;&comma; mathematics&comma; and program source code discussions&period;<&sol;p>&NewLine;<p>StarCoder2 advances the potential of future AI‑driven coding applications&comma; including text‑to‑code and text‑to‑workflow capabilities&period; With broader&comma; deeper programming training&comma; it provides repository context&comma; enabling accurate&comma; context‑aware predictions&period; These advancements serve seasoned software engineers and citizen developers alike&comma; accelerating business value and digital transformation&period;<&sol;p>&NewLine;<p>Users can fine‑tune the open‑access models with industry or organization‑specific data using open‑source tools such as NVIDIA NeMo or Hugging Face TRL&period;<&sol;p>&NewLine;<p>Organizations have already fine‑tuned the foundational StarCoder model to create specialized task‑specific capabilities for their businesses&period;<&sol;p>&NewLine;<p>ServiceNow’s text‑to‑code Now LLM was purpose‑built on a specialized version of the 15 billion‑parameter StarCoder LLM&comma; fine‑tuned and trained for ServiceNow workflow patterns&comma; use‑cases&comma; and processes&period; Hugging Face also used the model to create its StarChat assistant&period;<&sol;p>&NewLine;<p><b>BigCode Fosters Open Scientific Collaboration in AI<&sol;b><&sol;p>&NewLine;<p>BigCode represents an open scientific collaboration jointly led by Hugging Face and ServiceNow&period; Its mission centers on the responsible development of LLMs for code&period;<&sol;p>&NewLine;<p>The BigCode community actively participated in the technical aspects of the StarCoder2 project through working groups and task forces&comma; leveraging ServiceNow’s Fast LLM framework to train the 3 billion‑parameter model&comma; Hugging Face’s nanotron framework for the 7 billion‑parameter model&comma; and the end‑to‑end NVIDIA NeMo cloud‑native framework and <a href&equals;"https&colon;&sol;&sol;developer&period;nvidia&period;com&sol;blog&sol;nvidia-tensorrt-llm-supercharges-large-language-model-inference-on-nvidia-h100-gpus&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">NVIDIA TensorRT‑LLM<&sol;a> software to train and optimize the 15 billion‑parameter model&period;<&sol;p>&NewLine;<p>Fostering responsible innovation is at the core of BigCode’s purpose&comma; demonstrated through its open governance&comma; transparent supply chain&comma; use of open‑source software&comma; and the ability for developers to opt data out for training&period; StarCoder2 was built using responsibly sourced data under license from the digital commons of <a href&equals;"https&colon;&sol;&sol;www&period;softwareheritage&period;org&sol;2024&sol;02&sol;28&sol;responsible-ai-with-starcoder2&sol;" target&equals;"&lowbar;blank" rel&equals;"noopener">Software Heritage<&sol;a>&comma; hosted by <a href&equals;"https&colon;&sol;&sol;www&period;inria&period;fr&sol;en&sol;inria-ecosystem" target&equals;"&lowbar;blank" rel&equals;"noopener">Inria<&sol;a>&period;<&sol;p>&NewLine;<p>&OpenCurlyDoubleQuote;StarCoder2 is the first code generation AI model developed using the Software Heritage source code archive and built to align with our policies for responsible development of models for code&comma;&&num;8221&semi; stated Roberto Di Cosmo&comma; Director at Software Heritage&period; &&num;8220&semi;The collaboration of ServiceNow&comma; Hugging Face&comma; and NVIDIA exemplifies a shared commitment to ethical AI development&comma; advancing technology for the greater good&period;&&num;8221&semi;<&sol;p>&NewLine;<p>StarCoder2&comma; as with its predecessor&comma; will be made available under the BigCode Open RAIL‑M license&comma; allowing royalty‑free access and use&period; Furthermore&comma; the supporting code for the models resides on the BigCode project’s GitHub page&period;<&sol;p>&NewLine;<p>All StarCoder2 models will also be <a href&equals;"https&colon;&sol;&sol;huggingface&period;co&sol;bigcode" target&equals;"&lowbar;blank" rel&equals;"noopener">available for download<&sol;a> from Hugging Face and the StarCoder2 15B model is available on <a href&equals;"https&colon;&sol;&sol;catalog&period;ngc&period;nvidia&period;com&sol;orgs&sol;nvidia&sol;teams&sol;ai-foundation&sol;models&sol;starcoder2-15b" target&equals;"&lowbar;blank" rel&equals;"noopener">NVIDIA AI Foundation models<&sol;a> for developers to experiment with directly from their browser&comma; or through an API endpoint&period;<&sol;p>&NewLine;

Editor

Intuit Signs $100 Million Deal With OpenAI

MOUNTAIN VIEW - Intuit Inc., operator of the financial technology platform that makes Intuit TurboTax, Credit Karma, QuickBooks,…

2 days

Heavily Armed Drug Dealer Arrested Next to Sunnyvale School

The Santa Clara County District Attorney’s Office has charged a Sunnyvale apartment manager with possessing…

2 days

Wispr Scores $25 Million Series A Extension

SAN FRANCISCO -- Wispr, the voice-to-text AI that turns speech into clear, polished writing in every…

3 days

Numeric Dials Up $51 Million Series B

SAN FRANCISCO -- Numeric, an AI accounting automation platform, has raised a $51 million Series…

3 days

Apple Names 45 Finalists for App Store of the Year Awards

Apple has announced 45 finalists for this year’s App Store Awards, recognizing the best apps…

4 days

UC Reaches Agreement With Nurses, Strike Canceled

The University of California (UC) and the California Nurses Association (CNA) have reached a tentative…

6 days