NVIDIA • Mistral • Arc Institute • & Others… +167% / Year Both models from DeepMind (AlphaGo Zero & Master) Publication Date19 ChatGPT AI User + Subscriber + Revenue Growth Ramps = Hard to Match with performance in line with Western competitors 1/25: DeepSeek releases its R1 & R1- Zero open- source reasoning models 2/25: OpenAI releases GPT-4.5, Anthropic releases Claude manually curates the data, some models considered notable by some may not be included. A count of zero academic models does not mean that no notable models were produced by academic institutions in 2023

. . . . . . . . . . . . . . . 7 3.1.1 核心创新 1：含 R1-Zero 的中间推理模型 . . . . . . . 8 3.1.2 核心创新 2：通用强化学习 . . . . . . . . . . . . . . . 8 3.2 含 R1-Zero 的中间推理模型训练过程 . . . . . . . . . . . . . . 9 3.3 通用强化学习训练过程 R1 完整训练过程 训练起点。DeepSeek-R1 的训练起点是 DeepSeek-v3-Base，作为基础模型 进行训练，为后续的推理优化奠定基础。 3.1.1 核心创新 1：含 R1-Zero 的中间推理模型 如图7所示，推理导向的强化学习（Reasoning-Oriented Reinforcement Learn- ing）得到中间推理模型（Iterim reasoning model） 大模 型推理能力，开源纯强化学习推理模型 DeepSeek-R1-Zero R1-Zero 能生成高质量的推理数据，包括大量长链式思维（Chain-of-Thought, CoT）示例，用于支持后续的 SFT 阶段，如图7所示。更加详细介绍参考3.2节。 3.1.2 核心创新 2：通用强化学习 第一阶段 R1-Zero 虽然展现出惊人的推理能力提升，但是也出现了回复时 语言混合，非推理

Temperature 9 Top-K and top-P 10 Putting it all together 11 Prompting techniques 13 General prompting / zero shot 13 One-shot & few-shot 15 System, contextual and role prompting 18 System prompting 19 Role let’s dive into some examples of the most important prompting techniques. General prompting / zero shot A zero-shot5 prompt is the simplest type of prompt. It only provides a description of a task and some started with. This input could be anything: a question, a start of a story, or instructions. The name zero-shot stands for ’no examples’. Prompt Engineering February 2025 14 Let’s use Vertex AI Studio (for

as a proxy for FP32 number (not a downcast) • Quantized tensor is represented with a scale and a zero point http://on-demand.gputechconf.com/gtc/2017/presentation/s7310-8-bit-inference-with-tensorrt Operator fn (%input_data: Tensor[(2, 5), float32]) { qnn.quantize(%input_data, out_dtype="uint8", output_zero_point=127, output_scale=0.5f) } def @main(%input_data: Tensor[(2, 5), float32]) -> Tensor[(2, 5) Services, Inc. or its Affiliates. All rights reserved. Lowering of QNN Conv2D Operator For zero-centered zero point, the lowering will have just nn.conv2d fn (%data: Tensor[(1, 3, 2, 3), uint8], %weight:

internally by our engineers. It employs a 16-way zero-bubble pipeline parallelism (Qi et al., 2023), an 8-way expert parallelism (Lepikhin et al., 2021), and ZeRO-1 data parallelism (Rajbhandari et al., 2020) P. Qi, X. Wan, G. Huang, and M. Lin. Zero bubble pipeline parallelism. arXiv preprint arXiv:2401.10241, 2023. S. Rajbhandari, J. Rasley, O. Ruwase, and Y. He. Zero: Memory optimizations toward training $x-2 \ge 0$, so $x\ge2$, and $5 - x \ge 0$, so $x \le 5$. Also, the denominator cannot be equal to zero, so $5-x>0$, which gives $x<5$. Therefore, the domain of the expression is $\boxed{[2,5)}$. Final

例：课后作业 仔细思考政企、创业者必读 DeepSeek-R1是AI发展史上的重要里程碑 R1形成了新的AGI定律，加速了AGI发展 Alpha Zero时刻 • Alpha Go采用监督学习， 使用人类棋谱训练 • Alpha Zero采用强化学习， 自己跟自己对弈 ChatGPT时刻 • OpenAI ChatGPT大模型， 通过预训练方式，实现涌 现，理解人类语言和知识 • 诞生预训练Scaling

。 • 英伟达、微软、亚马逊等国际巨头纷纷接入DeepSeek。 DeepSeek R1引发全球关注 推理能力：核心突破，专项升级  推理能力 • 强化学习驱动：DeepSeek R1-Zero 是首个完全基于强化学习（RL） 训练的推理模型，无需任何监督微调（SFT）步骤，打破传统模型依 赖大量标注数据的惯例。DeepSeek-R1 采用强化学习作为核心训练 方法，显著提升了模型的推理能力和语言表达的可读性。 分高质量、结构化的数据。其作用是为模型提供一个良好的起 点，解决强化学习训练初期的不稳定问题，规范模型的输出格 式和推理链条，使其更符合人类可读性。 • 数据来源与特点：这些数据部分来源于清理后的R1-Zero 输出， 还包括人工后处理的长思维链（CoT）数据。其数量相对较少 但质量高，经过精心设计，具有良好的可读性和结构化特点。 • 对模型训练的影响：冷启动数据为模型训练奠定了坚实的基础， 使

Scaling the loss using S 了 Backward propagation in MP N 放gradients( Y ) Unscaled gradients Zero gr: adients Apply gradients 计算平台事业部 COMPUTING PLATFORM COMPUTING PLATFORM INT8 Inference on PAI-