译 GPT-4 官方技术报告( 七 ) _模型

我们感谢微软的合作，特别是微软Azure对模型训练的基础设施设计和管理的支持，以及微软Bing团队和微软安全团队关于安全部署的合作。
我们感谢我们的专家对抗性测试人员和渗透团队人员，他们在开发的早期阶段帮助测试我们的模型，并为我们的风险评估以及系统卡提供信息。参与这个渗透团队过程并不代表对的部署计划或的政策的认可：,Duba, Cèsar Ferri,Frase, Gavin , Jake J. Hecla, Dan , Jose -, Alice , Rajiv W. Jain, Boru Gollo ,Kahn, Dan , Sara , Noam Kolt,, Eric ,J. Lohn,, Sam ,, Anna Mills, Yael Moros, Jimin Mun, Aviv , Roya , Yifan Peng, Ciel Qi, Alex , Paul R?ttger,Sap, Wout ,Shih,,, Bryan West,D. White, Anna,, Lexin Zhou, Xuhui Zhou 。
我们感谢我们在和斯坦福CodeX的合作者进行模拟律师考试：P.(/ CodeX), D. Katz ( CodeX), M.( CodeX), S. Gao () 。
在这项工作中，GPT-4在措辞、格式和风格方面得到了帮助。
参考文献
[1] ,,,,.,,,Shyam,,, et al.are few-shot .in, 33:1877–1901, 2020.
[2],,, Elena ,Cai, Eliza , Diego de Las Casas, Lisa Anne ,Welbl, Aidan Clark, et al.- large. arXivarXiv:2203.15556, 2022.
[3],, Jacob ,Bosma,, Adam , Paul , Hyung Won Chung,,, et al. PaLM:with . arXivarXiv:2204.02311, 2022.
[4] Jack W Rae,,Cai, Katie ,,Song, John , Sarah , Roman Ring,Young, et al.: ,&from. arXivarXiv:2112.11446, 2021.rXiv:1810.04805*, 2018.
[8] Colin , Noam , Adam ,Lee,,, Yanqi Zhou, Wei Li, and Peter J Liu.theofwith atext-to-text . arXivarXiv:1910.10683, 2019.
[9] NoamandStern. :rates withcost. arXivarXiv:1804.04235, 2018.
[10] Jimmy Lei Ba, Jamie Ryan Kiros, andE. . Layer . arXivarXiv:1607.06450, 2016.
[11] Jason Wei,Wang, Dale ,Bosma, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-in large. , 2022.
[12]Huang,Shane Gu, Le Hou,Wu,Wang,Yu, andHan. Largecan self-. arXivarXiv:2210.11610, 2022.
[13],Shane Gu,Reid,, and. Largeare zero-shot . arXivarXiv:2205.11916, 2022.
[14] Jared , Sam , Tom , Tom B. Brown,Chess, Rewon Child, Scott Gray, Alec ,Wu, and Dario .laws for. arXivarXiv:2001.08361, 2020.
[15] Tom , Jared , Mor Katz, Mark Chen,Hesse, Jacob ,Jun, Tom B. Brown,, Scott Gray, et al.laws for. arXivarXiv:2010.14701, 2020.
[16] Greg Yang,J. Hu, Igor ,Sidor,Liu, David Farhi, Nick Ryder, Jakub ,Chen, andGao.V:largevia zero-shot. arXivarXiv:2203.03466, 2022.
[17] Noam ,,, Andy Davis, Quoc Le,, and Jeff Dean.large: The -gated -of- layer. arXivarXiv:1701.06538, 2017.
[18]Zoph, Irwan Bello,Kumar, Nan Du,Huang, Jeff Dean, Noam , andFedus. ST-MoE:and. arXivarXiv:2202.08906, 2022.
[19] Jason Wei, Yi Tay, Rishi , Colin ,Zoph,, Dani ,Bosma, Denny Zhou,, et al.of large. TMLR, 2022.
[20],Gouws, Oriol , Jakob , and. Uni-. Inon, 2019. URL
[21]Su, Yu Lu,Pan, Ahmed , Bo Wen, andLiu. :with. arXivarXiv:2104.09864, 2021.
[22] Jean- , Jeff ,Luc,Miech, Iain Barr, Yana , Karel Lenc,,,, et al. : amodel for few-shot . Inin.
[23] Xi Chen, Xiao Wang,, AJ , Piotr ,Salz,, Adam , Basil , Lucas Beyer, et al. PaLI: A --image model. arXivarXiv:2209.06794, 2022.
[24] Ben Wang and Aran . GPT-J-6B: A 6model, 2021.

文章插图
[25] Sid Black, Leo Gao, Phil Wang,Leahy, and. GPT-Neo: Large scalewith mesh-. If you use this ,cite it using these , 58, 2021.
[26] Teven Le Scao,Fan,Akiki, Ellie ,Ilic ?,, Roman é,Sasha , Fran?ois Yvon,Gallé, et al. Bloom: A 176B- open-model. arXivarXiv:2211.05100, 2022.
[27] Susan Zhang,, Naman Goyal, Mikel , Moya Chen,Chen,Dewan, Mona Diab, Xian Li, XiLin, et al. OPT: Open pre-. arXivarXiv:2205.01068, 2022.
[28] Hugo ,,,, Marie-Anne , Timo- thée ,Rozière, Naman Goyal, Eric ,Azhar, et al. LLaMA: Open and. arXivarXiv:2302.13971, 2023.
[29] Alec , Rafal Jó, and Ilya .toand - ing . arXivarXiv:1704.01444, 2017.
[30]and. Cross-model . arXivarXiv:1901.07291, 2019.
[31] Tri Dao,Y. Fu,Ermon, Atri Rudra, andRé. : Fast and - exactwith io-. arXivarXiv:2205.14135, 2022.