第一项始终是图钉的来源。总体可信度是各份资料对上方所用开始和结束时间支持程度的加权平均;资料每比最新的一份旧 180 天,权重减半。
谷歌的文章《Gemma 2 现已面向研究人员和开发者开放》发布于2024年6月27日(署名为 Clement Farabet 和 Tris Warkentin),Maginative[5] 当天即作了报道;Hugging Face[4][6] 仓库于6月24日在禁运期内创建,技术报告随后于2024年7月31日发布在 arXiv[2] 上。
Gemma 2 technical report (v1 31 July 2024): 2B to 27B models with alternating local-global attention and grouped-query attention, 2B and 9B trained by knowledge distillation, 27B on 13T tokens and 9B on 8T; LMSYS Chatbot Arena Elo 1218 for the 27B and 1187 for the 9B.[1]
OpenRouter listing (read from its models API): Google: Gemma 2 27B, added 13 July 2024, 8,192-token context, $0.65 per million input and output tokens.
Model card (gated): text-to-text decoder-only English model, 13 trillion training tokens for the 27B and 8 trillion for the 9B, Gemma license, pretrained benchmark table with MMLU 75.2, HellaSwag 86.4, GSM8K 74.0 and HumanEval 51.8 for the 27B.[1]
Maginative (Chris McKay), 27 June 2024: Gemma 2 9B and 27B released, 27B competitive with proprietary models more than twice its size and approaching Llama 3 70B and Claude 3 Sonnet, single TPU host, A100 80GB or H100, commercially friendly license, available in Google AI Studio and on Kaggle and Hugging Face[4][6].
有遗漏或错误吗?用你自己的话说明:能佐证此图钉的链接、不同的开始或结束日期及理由,或缺失、有误的信息。AI 会对照此图钉的来源进行核实,搜索更好的来源,并添加任何支持你说法的页面。图钉自身的来源仍然最重要。AI 也会查看图片:显示的是别的东西或显示效果差的图片会被移到后面或替换。
原标题: Gemma 2 9B and 27B Released
谷歌发布博文中的性能图表;“-”表示未报告[1].
| 基准测试(指标) | Gemma 2 9B | Gemma 2 27B | Llama 3 8B | Llama 3 70B | Grok-1 314B |
|---|---|---|---|---|---|
| MMLU(5-shot,top-1) | 71.3 | 75.2 | 66.6 | 79.5 | 73.0 |
| BBH (3-shot, CoT) | 68.2 | 74.9 | 61.1 | 81.3 | - |
| HellaSwag (10-shot) | 81.9 | 86.4 | 82 | - | - |
| GSM8K (5-shot, maj@1) | 68.6 | 74.0 | 45.7 | - | 62.9(8-shot) |
| MATH (4-shot) | 36.6 | 42.3 | - | - | 23.9 |
| HumanEval (pass@1) | 40.2 | 51.8 | - | - | 63.2(0-shot) |