Skip to main content

> ML_LITERATURE // GEMMA-TEAM-2024-GEMMA-OPEN-MODELS_v1.0

Gemma: Open Models Based on Gemini Research and Technology

Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Aida Nematzadeh, Sebastian Borgeaud, Emanuele Bugliarello, Albin Cassirer, Pier Giuseppe Sessa, Abe Fries, Alecs Lybus, Alex Verga, Alex Fabrikant, Alexander B. Wang, Alexander Goldin, Alexandru Damian, Alireza Ghaffarkhah, Amit Varma, Andreas Fidjeland, Andrew Dai, Ankur Parikh, Antoine Miech, Arman Bitton, Arthur Douillard, Arthur Guez, Asha Sreedhar, Ashok Kumar, Balaji Venkatraman, Boaz Barak, Bobak Shahriari, Canfer Akbulut, Carlos A. Gomez-Uribe, Clemens Meyer, Colin Gaffney, Corentin Tallec, Craig Boutilier, Danila Sinopalnikov, David A. D'Ambrosio, David Reitter, Diana Mincu, Dmitry Lepikhin, Douglas Eck, Dror Marcus, Elena Buchatskaya, Eleni Triantafillou, Ellis Whitehead, Fabian Mentzer, Fady Medhat, Gage Hills, Geoff Brown, George Chen, Gaurav Mishra, Gilles Baechler, Gus Lopez, Haggai Nativ, Hanna Klimczak-Plucińska, Hao Zhang, Hao-Tse Cheng, Hemanth Makkapati, Hillary Richardson, Hongmin Xu, Houssam Zenati, Ian Ballantyne, Ilya Kostrikov, Izzeddin Gur, Jack W. Rae, James Besley, James Lee-Thorp, James Manyika, Jean-Baptiste Lespiau, Jeff Stanway, Jie Ren, Jing Li, Jingfei Du, John Aslanides, John Dillon, John Guilyardi, Jonathan Lai, Jonathan Mallinson, Jonathan Shen, Jordan Grimstad, Joseph Joshi, Jovana Mitrovic, Kallolini Bora, Kavya Kopparapu, Keith Anderson, Kelvin Xu, Kevin Villela, Kiran Vodrahalli, Kishore Papineni, Kristian Lum, Ksenia Konyushkova, Kyle Levin, Lalida Sritanyaratana, Lee Carveth, Leila Wehbe, Lê Nguyên, Lili Guan, Lisa Lee, Lucas Dixon, Lukas Zilka, Maciej Mikuła, Mariya Kovalenko, Mark Heffernan, Markus Kunesch, Martin Benda, Mateo Wirth, Mathieu Ravaut, Matthew Mauger, Matthew Lamm, Maxim Krikun, Michael Sharman, Michael Voronov, Michelle Casbon, Mike Schaekermann, Mona Chitnis, Monica Caraway, Moustafa Alzantot, Natalia Ponomareva, Naveen Kumar, Nemanja Rakićević, Noam Shazeer, Nora Kassner, Norbert Kalb, Oliver Woodman, Olivier Bousquet, Pablo Sprechmann, Pengcheng Yin, Peter Humphreys, Peter Liu, Petko Georgiev, Pradeep Shenoy, Prateek Jain, Pushmeet Kohli, Rachel Hornung, Radu Soricut, Raja Patel, Rajat Monga, Raveesh Sharma, Reza Rokni, Rishabh Agarwal, Ryan Mullins, Sahar Noorzad, Samy Bengio, Satinder Baveja, Sebastian Goodman, Sebastien M. R. Arnold, Seymour Knowles-Barley, Shachar Mirkin, Shashi Narayan, Shilpa Agrawal, Shravan Nayak, Shuai Liu, Siddhartha Brahma, Siva Reddy, Soroush Radpour, Srihari Thirumalai, Srinivasan Venkatachary, Stella Biderman, Susan Chan, Suzan Schwartz, Swaroop Mishra, Tomer Levinboim, Tris Warkentin, Wei-Ning Hsu, Willem de Groot, William Ferguson, Xinyun Chen, Yana Hasson, Yarden Sarig, Yossi Matias, Yuancheng Chu, Yuxin Bai, Zack Ontiveros, Zhiyu Chen, Zoubin Ghahramani, Demis Hassabis, Koray Kavukcuoglu, Jeffrey Dean, Oriol Vinyals · arXiv preprint (2024)

seminal-architecture2024industry-standardthirdPartyReproduced

Temel Katkı (Principal Contribution)

Google Gemini mimari yeniliklerini (multi-query attention, GeGLU aktivasyonları, normalleştirici yerleşimi) 2B ve 7B açık ağırlıklı modellere aktardı.

Operasyonel ve Mühendislik Uygunluğu

task-text-generation, task-question-answering için mimari kararları, hizalama stratejisini ve sunum altyapısını doğrudan yönlendirir.

Temel Varsayımlar (Assumptions)

  • Ampirik dağılım düzenliliği geçerlidir ve hedef alan ön eğitim dilsel/görsel desteğine uygundur

Kısıtlar ve Sınırlamalar (Limitations)

  • Kaynak ölçekleme, çıkarım bellek gereksinimleri ve hizalama sağlamlığı model boyutuna ve donanım topolojisine göre değişir

Bağlantılı Algoritmalar, Mimariler ve Kütüphaneler

İlgili Algoritmalar:
İlgili Mimariler:
Uygulayıcı Kütüphaneler: