“They even show plots that clearly demonstrate they settled on the metric because it's the one where GPT-4 results remain they same, but Gemini wins.”
Gemini: A Family of Highly Capable Multimodal Models
介绍 Ultra、Pro、Nano 三档多模态 Gemini 模型及其评测结果。
报告介绍 Gemini 多模态模型家族,覆盖图像、音频、视频和文本理解,分 Ultra、Pro、Nano 三种规模,分别面向复杂推理和端侧受限场景。最强的 Gemini Ultra 在所评测的 32 个基准中的 30 个上刷新最优,并称是首个在 MMLU 上达到人类专家水平的模型,在所评测的 20 个多模态基准上全部取得最优。报告还讨论了后训练以及通过 Gemini、Google AI Studio、Cloud Vertex AI 等服务负责任地部署。
评论者把它当技术报告引用来核对细节,如视觉 token 的处理、Nano 的参数规模(1.8B 和 3.25B)和预训练数据构成,也有人批评其评测指标是挑出来让 Gemini 胜出的。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“It only takes looking at the technical report for the Gemini models to see that you're completely wrong.”
“From the Gemini report Gemini”
“That feels like it runs counter to this statement from the Gemini 1.0 technical report[0]:”
“According to the paper[1], Gemini Nano-1 is 1.8B, and (Nano-2) is 3.25B parameters.”
“Our pre-training dataset uses data from web documents, books, and code, and includes image, audio, and video data.”
- Gemini: A Family of Highly Capable Multimodal Models ↗lysozyme · 2023-12-20 · 2 分
- Gemini: A Family of Highly Capable Multimodal Models ↗Multiset · 2023-12-20 · 2 分