Author
Arjun Guha (Roblox + Northeastern University), Raymond Li (ServiceNow), Loubna Ben Allal (HuggingFace), Yangtian Zi (Northeastern University), Niklas Muennighoff (HuggingFace), Denis Kocetkov (ServiceNow), Chenghao Mou (Independent), Marc Marone (Johns Hopkins University), Christopher Akiki (Leipzig University + ScaDS.AI), Jia Li (Independent), Jenny Chim (Queen Mary University of London), Qian Liu (Sea AI Lab), Evgenii Zheltonozhskii (Technion – Israel Institute of Technology), Terry Yue Zhuo (Monash University + CSIRO’s Data61), Thomas Wang (HuggingFace), Olivier Dehaene (HuggingFace), Mishig Davaadorj (HuggingFace), Joel Lamy-Poirier (ServiceNow), João Monteiro (ServiceNow), Oleh Shliazhko (ServiceNow), Nicolas Gontier (ServiceNow), Nicholas Meade (Mila + McGill University), Armel Zebaze (HuggingFace), Ming-Ho Yee (Northeastern University), Logesh Kumar Umapathi (Saama AI Research Lab), Jian Zhu (University of British Columbia), Benjamin Lipkin (MIT), Muhtasham Oblokulov (Technical University of Munich), Zhiruo Wang (Carnegie Mellon University), Rudra Murthy (IBM Research), Jason Stillerman (University of Vermont), Siva Sankalp Patel (IBM Research), Dmitry Abulkhanov (Independent), Marco Zocca (UnfoldML), Manan Dey (SAP), Zhihan Zhang (University of Notre Dame), Nour Fahmy (Columbia University), Urvashi Bhattacharyya (Discover Dollar Pvt Ltd), Wenhao Yu (University of Notre Dame), Swayam Singh (University of Allahabad), Sasha Luccioni (HuggingFace), Paulo Villegas (Telefonica I+D), Maxim Kunakov (Toloka), Fedor Zhdanov (Toloka), Manuel Romero (Independent), Tony Lee (Stanford University), Nadav Timor (Weizmann Institute of Science), Jennifer Ding (The Alan Turing Institute), Claire Schlesinger (Northeastern University), Hailey Schoelkopf (Eleuther AI), Jan Ebert (Forschungszentrum Jülich), Tri Dao (Stanford University), Mayank Mishra (IBM Research), Alex Gu (MIT), Jennifer Robinson (ServiceNow), Carolyn Jane Anderson (Wellesley College), Brendan Dolan-Gavitt (NYU), Danish Contractor (Independent), Siva Reddy (ServiceNow + Mila), Daniel Fried (Carnegie Mellon University), Dzmitry Bahdanau (ServiceNow), Yacine Jernite (HuggingFace), Carlos Muñoz Ferrandis (HuggingFace), Sean Hughes (ServiceNow), Thomas Wolf (HuggingFace), Leandro von Werra (HuggingFace), Harm de Vries (ServiceNow)
Venue
Transactions on Machine Learning Research (TMLR) 2023
Abstract
The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.
