AWS AIF-C01 · AI 真题库（452 题） · 第 03 份

Amazon SageMaker Clarify is a tool that helps detect bias in machine learning models and provides explainability for model predictions. It allows users to understand the factors influencing a model's predictions and assess whether the model is fair across different demographic groups. This is exactly what the company needs for detecting bias and explaining model decisions, especially for high-stakes applications like loan approvals. -Bias detection: SageMaker Clarify can analyze the training data and the model's predictions to identify and mitigate bias. -Explainability: It provides features for explaining the predictions made by the model, which helps users understand the reasons behind the model'sdecisions.

亚马逊SageMakerClarify是一款有助于检测机器学习模型中的偏差并为模型预测提供解释性的工具。它能 让用户了解影响模型预测的因素，并评估模型在不同人口群体中的公平性。这正是该公司检测偏差和解释模 型决策所需要的，尤其是在贷款审批等高风险应用中。 ，偏差检测：SageMakerClarify可以分析训练数据和模型的预测结果，以识别并减轻偏差。 可解释性：它提供了用于解释模型所做预测的功能，这有助于用户理解模型决策背后的原因。 .

BERTScore is a metric specifically designed for evaluating the quality of text generated by models, particularly in tasks like text summarization and machine translation. It uses contextual embeddings from a pre-trained BERT model to compare generated text with reference text at the word level, making it well- suited for evaluating the accuracy of generative text summarization models. BERTScore assesses the semantic similarity between the generated text and the reference text, providing a more nuanced evaluation compared to traditional methods that focus on exact matches.

BERTScore是一种专门用于评估模型生成文本质量的指标，尤其适用于文本摘要和机器翻译等任务。它利 用预训练的BERT模型的上下文嵌入，在单词层面将生成的文本与参考文本进行比较，因此非常适合评估生 成式文本摘要模型的准确性。 BERTScore评估生成文本与参考文本之间的语义相似度，与侧重于精确匹配的传统方法相比，它能提供更 细致的评价。

K-nearest neighbors (k-NN) is a supervised learning algorithm used for classification tasks. It works by classifying a data point based on how its features (such as petal length, petal width, sepal length, and sepal width in this case) are similar to the data points in the training set. For this problem, where the task is to classify flowers based on their features, k-NN is an appropriate choice. k-NN is simple and effective for classification problems where the decision boundary is not necessarily linear, as it looks at the closest neighbors to make predictions.

K近邻（k-NN）是一种用于分类任务的监督学习算法。它通过将数据点的特征（例如本例中的花瓣长度、 花瓣宽度、萼片长度和萼片宽度）与训练集中的数据点进行比较来对其进行分类。对于此问题，即根据特征 对花进行分类，k-NN是一个合适的选择。 k近邻算法简单有效，适用于决策边界不一定为线性的分类问题，因为它通过查看最近的邻近点来进行 预测。

AWS Key Management Service (AWS KMS) is a fully managed service that allows you to create and control encryption keys used to encrypt your data. It is ideal for scenarios like this, where a company wants to use a custom, company-managed encryption key to protect model artifacts. AWS KMS allows you to securely manage keys for encrypting and decrypting data, and it integrates with various AWS services, including Amazon Bedrock. Aws KMs provides centralized key management, and it is designed for use cases that involve encrypting both data and artifacts, such as model customization jobs.

AWS密钥管理服务（AWSKMS）是一项完全托管的服务，可让您创建和控制用于加密数据的加密密钥。 对于公司希望使用自定义的、由公司管理的加密密钥来保护模型资产的场景，它非常理想。AWSKMS允 许您安全地管理用于加密和解密数据的密钥，并且它与包括AmazonBedrock在内的多种AWS服务集成。 AWSKMS提供集中式密钥管理，专为涉及加密数据和工件（例如模型定制任务）的用例而设计。

Text generation is the feature of large language models (LLMs) that enables them to produce text based on a given prompt or input. In this scenario, the input would be natural language code comments, and the LLM would generate code based on those comments. This process involves understanding the context of the comment and generating corresponding code, which is the core functionality of text generation.

文本生成是大型语言模型（LLM）的一项功能，使它们能够根据给定的提示或输入生成文本。在这种情 况下，输入将是自然语言的代码注释，而LLM将根据这些注释生成代码。这一过程涉及理解注释的上 下文并生成相应的代码，这是文本生成的核心功能。

ROUGE is a set of metrics that evaluates the quality of summaries by comparing the overlap of n-grams, word sequences, and word pairs between the model output and reference examples. Since the company is working on a language model that improves the coherence of text and wants the output to resemble the provided examples (which are more readable versions of the original text), ROUGE is the most appropriate metric. It assesses how closely the generated text matches the reference text in terms of content and readability. RoUGE score is commonly used to evaluate the performance of models in tasks like summarization, where the goal is to ensure the generated text aligns closely with human-provided examples.

ROUGE是一组评估摘要质量的指标，通过比较模型输出与参考示例之间的n元语法、词序列和词对的 重叠情况来实现。由于该公司正在开发一种能提高文本连贯性的语言模型，并希望其输出与提供的示例 （即原始文本的更易读版本）相似，因此ROUGE是最合适的评估指标。它能评估生成的文本在内容和 可读性方面与参考文本的接近程度。 ROUGE分数通常用于评估诸如摘要生成等任务中模型的表现，其目的是确保生成的文本与人工提 供的示例高度一致。

Negative prompts are used to explicitly instruct the model about what to avoid or not generate. By providing the model with specific guidance on what is not desired (e.g., by including terms or concepts that should not appear in the image), the model can better focus on generating relevant and related content. This technique can help reduce unrelated or irrelevant images by constraining the model's creative generation process.

负向提示用于明确告知模型应避免或不生成的内容。通过向模型提供具体指导，说明哪些内容是不希望 出现的（例如，包含不应出现在图像中的术语或概念），模型能够更好地专注于生成相关且恰当的内容。 这种技术有助于减少不相关或无关的图像，从而限制模型的创意生成过程。

To generate concise, feature-specific descriptions for each product, the company should create prompts tailored to specific product categories. By highlighting the key features of each product category in the prompt, the model can focus on generating descriptions that are relevant and aligned with the unique attributes of each product. Additionally, specifying the desired output format and length ensures that the responses meet the company's requirements for conciseness and clarity. Tailored prompts help ensure the model generates relevant and accurate descriptions by focusing on the most important features of each product category. Desired output format and length ensure the responses are consistent and concise, as required.

为了为每款产品生成简洁、突出特定功能的描述，公司应当针对不同的产品类别制定专门的提示。通 过在提示中突出每个产品类别的关键特性，模型能够专注于生成与每款产品的独特属性相关且一致的 描述。此外，明确期望的输出格式和长度，能够确保 回复符合公司对简洁明了的要求。 定制化的提示有助于确保模型生成的相关描述准确无误，重点突出每个产品类别的最重要特征。 期望的输出格式和长度确保响应一致且简洁，符合要求。

The issue described is a common case of overfitting, where the model performs well on the training data but fails to generalize to new, unseen data. This suggests that the model is too complex and has learned to memorize the training data rather than identifying generalizable patterns. Increasing the regularization parameter helps to reduce model complexity by penalizing large weights, thereby encouraging simpler models that are less likely to overfit. This will improve the model's ability to generalize to new data, potentially improving performance on unseen customer data.

所描述的问题是常见的过拟合情况，即模型在训练数据上表现良好，但在新的、未见过的数据上却无法泛 化。这表明模型过于复杂，已经学会记住训练数据，而不是识别可泛化的模式。 增加正则化参数有助于通过惩罚较大的权重来降低模型的复杂度，从而鼓励生成更简单的模型，使其不太 可能过拟合。这将提高模型对新数据的泛化能力，从而有可能提升其在未见过的客户数据上的表现。

Amazon Aurora PostgresQL is a relational database service that supports a wide range of applications, including those that require vector-based operations. Specifically, Aurora PostgreSQL can be extended with vector search capabilities using extensions like pgvector. This extension allows you to store, index, and query vector embeddings from generative Al models, making it well-suited for the company's needs to store and query embeddings as vectors. pgvector is an extension for PostgresQL that provides efficient similarity search for vector data, which is ideal for storing and querying embeddings.

AmazonAuroraPostgresQL是一种关系型数据库服务，支持广泛的应用程序，包括那些需要基于向量的操 作的应用程序。具体而言，Aurora PostgreSQL可以通过诸如pgvector这样的扩展来增强向量搜索功能。此 扩展允许您存储、索引和查询来自生成式AI模型的向量嵌入，使其非常适合该公司存储和查询向量嵌入的 需求。 Pgyector是PostgreSQL的一个扩展，它为向量数据提供了高效的相似性搜索功能，非常适合用于存储和查 简嵌入式数据。一限，一主联的度

Model complexity plays a significant role in the explainability of an Al solution. Simpler models tend to be more explainable because it is easier to understand how they make decisions. For example, linear regression or decision trees are generally easier to interpret compared to more complex models like deep neural networks. In contrast, highly complex models, such as deep learning models, may provide very accurate results but are often considered "black boxes," meaning their decision-making process is not easily interpretable. This lack of transparency makes it difficult to explain their decisions in a way that satisfies regulatory requirements or customer understanding.

模型的复杂程度在人工智能解决方案的可解释性方面起着重要作用。较简单的模型往往更易于解释， 因为理解它们如何做出决策相对容易。例如，线性回归或决策树通常比深度神经网络等更复杂的模型 更容易解读。 相比之下，高度复杂的模型，例如深度学习模型，可能会提供非常准确的结果，但通常被视为“黑箱”， 这意味着其决策过程难以解释。这种缺乏透明度使得很难以满足监管要求或让客户理解的方式解释其决 策。

Amazon Bedrock provides access to a variety of large language models (LLMs), which are well-suited for tasks such as text summarization. By using LLMs, the pharmaceutical company can automatically generate concise and coherent summaries of user reviews for each medication. These models are designed to understand and process large amounts of text, making them ideal for summarizing user reviews in a clear and efficient manner. Amazon Bedrock allows the company to utilize LLMs that can generate summaries, which is exactly what is needed in this case.

亚马逊Bedrock提供了多种大型语言模型（LLM）的访问权限，这些模型非常适合诸如文本摘要之类 的任务。通过使用这些语言模型，制药公司能够自动为每种药物生成简洁且连贯的用户评论摘要。这 些模型旨在理解和处理大量文本，因此非常适合以清晰高效的方式总结用户评论。 亚马逊Bedrock使该公司能够利用能够生成摘要的大型语言模型，而这正是当前所需。

A logistic regression model is a simple, interpretable machine learning model that allows users to adjust the weights of different variables based on domain knowledge and expertise. Since logistic regression is based on a linear relationship between the input features and the predicted outcome, each feature has a corresponding weight (coefficient) that can be easily viewed and adjusted. This makes it a suitable choice for a lead prioritization application where domain experts need the ability to modify the model based on their knowledge.

逻辑回归模型是一种简单且易于解释的机器学习模型，它允许用户根据领域知识和专业经验调整不同变 量的权重。由于逻辑回归基于输入特征与预测结果之间的线性关系，每个特征都有一个对应的权重（系 数），这些权重可以轻松查看和调整。这使得它成为一种适用于潜在客户优先级排序应用的合适选择， 在这种应用中，领域专家需要能够根据自已的知识来修改模型。

The typical sequence of steps in building an ML application involves: 1. Defining the business goal and framing the ML problem - This step involves understanding the business need and determining how ML can address it. 2. Developing the model - This involves selecting the appropriate algorithm, training the model, and evaluating it. 3. Deploying the model - Once the model is trained, it is deployed to production to serve predictions. 4. Monitoring the model - After deployment, the model is monitored to ensure it performs well over time and to detect any issues like model drift.

构建机器学习应用程序的典型步骤包括： 1.明确业务目标并界定机器学习问题一一此步骤包括理解业务需求，并确定机器学习如何能够解决这些 尚题。 2.开发模型一一这包括选择合适的算法、训练模型以及对其进行评估。 3.部署模型-模型训练完成后，将其部署到生产环境中以提供预测服务。 4.模型监控-部署后，会对模型进行监控，以确保其长期表现良好，并检测诸如模型漂移之类的问题。

To determine if a foundation model (FM) effectively meets business objectives, the key is to assess how well the model aligns with the specific use cases or business goals it is intended to address. This involves evaluating whether the model can provide meaningful and accurate outputs relevant to the business's needs, ensuring that it solves the real-world problems the company aims to tackle. Aligning with specific use cases ensures that the model's capabilities are tailored to the tasks at hand, such as improving customer support, enhancing decision-making, or automating certain processes.

要确定基础模型（FM）是否有效地满足业务目标，关键在于评估该模型与它旨在解决的具体用例或业务目 标的契合程度。这包括评估模型能否提供与业务需求相关的有意义且准确的输出，确保其能够解决公司想 要应对的实际问题。 与具体的使用场景保持一致，可确保模型的能力符合手头的任务，例如改进客户支持、增强决策能力或自 动化某些流程。

In supervised learning, the model is trained using a labeled dataset, where each image (input) has a corresponding label (the type of animal, in this case). Since the company already has a large dataset of labeled images, supervised learning is the most appropriate approach. The model learns to classify the images based on the features and labels in the training data.

在监督学习中，模型是通过使用有标签的数据集进行训练的，其中每张图像（输入）都有一个对应的标 签（在这种情况下是动物的种类）。由于该公司已经拥有大量有标签的图像数据集，因此监督学习是最 合适的方法。模型根据训练数据中的特征和标签来学习对图像进行分类。 问题117--.--- 正确答案：C 在机器学习生命周期的数据收集阶段，确定合规性和监管要求最为关键。在此阶段，组织必须确保用于训练 模型的数据符合法律和监管标准，例如数据隐私法（如《通用数据保护条例》、《健康保险流通与责任法 案》）、行业特定法规以及道德考量。组织还必须核实其使用数据已获得适当授权，并且数据中不存在任何 偏见或违反任何法规的情况。

The data collection phase of the ML lifecycle is the most relevant for determining compliance and regulatory requirements. During this phase, organizations must ensure that the data being collected and used for training the model complies with legal and regulatory standards, such as data privacy laws (e.g., GDPR, HIPAA), industry-specific regulations, and ethical considerations. The organization must also verify that they have the proper consent to use the data and that the data does not contain any biases or violate any regulations.

To continuously improve the accuracy of the machine learning model, Amazon SageMaker is the appropriate solution. SageMaker allows for efficient model training, deployment, and iteration. The company can use SageMaker to retrain the model regularly with newer data to account for changes in customer behavior, market trends, and other dynamic factors that may affect food waste and sales revenue. Iterating with newer data helps improve the model's performance over time, ensuring it remains accurate and relevant.

为了持续提高机器学习模型的准确性，业马逊SageMaker是合适的解决方案。SageMaker能够实现高效的 模型训练、部署和选代。该公司可以使用SageMaker定期用更新的数据重新训练模型，以适应客户行为、 市场趋势以及其他可能影响食品浪费和销售收入的动态因素的变化。通过使用更新的数据进行选代，有助 于随着时间的推移提高模型的性能，确保其准确性和相关性。

Amazon SageMaker provides a fully managed service for deploying machine learning models at scale without needing to manage servers or infrastructure. When you deploy a model using Amazon SageMaker endpoints, it handles all the infrastructure, scaling, and maintenance aspects. This allows the company to focus solely on making predictions and integrating the model into their application, without worrying about managing servers or clusters. Amazon SageMaker simplifies the deployment of ML models by providing scalable, secure, and serverless endpoints, which are ideal for serving predictions in real-time.

亚马逊SageMaker提供了一项完全托管的服务，可大规模部署机器学习模型，无需管理服务器或基础设施。 当您使用业马逊SageMaker端点部署模型时，它会处理所有基础设施、扩展和维护方面的工作。这使得公 司能够专注于进行预测并将模型集成到其应用程序中，而无需担心管理服务器或集群。 业马逊SageMaker通过提供可扩展、安全且无服务器的端点来简化机器学习模型的部署，这些端 点非常适合实时提供预测服务。

Amazon Bedrock is a service that allows you to build Al-powered applications using foundation models (FMs) without managing infrastructure. Agents for Amazon Bedrock allows you to create intelligent agents that can help users interact with data and perform tasks such as checking open customer claims, identifying claim details, and accessing relevant documents. By using Amazon Bedrock knowledge bases, the Al agent can access relevant information (e.g., details about claims, documents, or customer data) in real time, and provide precise, context-driven answers based on the claim-related data stored in the knowledge base. - Agents for Amazon Bedrock can help automate interactions by understanding and responding to natural language queries, making it a good fit for this use case. : Amazon Bedrock knowledge bases enable the system to access stored information, documents, and details for specific claims efficiently.

亚马逊Bedrock是一项服务，可让您使用基础模型（FMs）构建AI驱动的应用程序，而无需管理基础设 施。亚马逊Bedrock代理允许您创建智能代理，帮助用户与数据交互并执行诸如检查未决客户索赔、识 别索赔详情以及访问相关文档等任务。 通过使用亚马逊Bedrock知识库，AI代理可以访问相关信息（例如，详情 能够实时处理有关索赔、文件或客户数据的查询，并根据知识库中存储的与索赔相关的数据提供精确、 基于上下文的答案。 ·亚马逊Bedrock的代理能够通过理解并回应自然语言查询来帮助实现交互自动化，因此非常适合这 种用例。 亚马逊Bedrock知识库使系统能够高效地访问存储的信息、文档以及特定索赔的详细资料。

Computer vision is a field of Al that enables machines to interpret and understand visual information from the world, such as images and videos. In the case of the manufacturing company, the Al is used to inspect products for damages or defects, which involves analyzing visual data (e.g., product images or videos). This is a classic application of computer vision, where the Al system identifies and classifies objects or defects within images.

计算机视觉是人工智能的一个领域，它能让机器解读和理解来自世界的视觉信息，比如图像和视频。在这家 制造公司，人工智能被用于检查产品是否存在损坏或缺陷，这涉及对视觉数据（例如产品图像或视频）进行 分析。这是计算机视觉的一个典型应用，其中人工智能系统能够识别并分类图像中的物体或缺陷。

Amazon SageMaker provides a fully managed environment for building, training, and deploying machine learning models, including automatic model tuning. Specifically, SageMaker includes a feature called Automatic Model Tuning (or Hyperparameter Optimization), which automates the process of finding the best hyperparameters for your machine learning model. This is essential when you want to optimize the model's performance without manual intervention. Amazon SageMaker allows you to automate the training process and hyperparameter tuning, which aligns perfectly with the company's need for fully automated model tuning.

亚马逊SageMaker提供了一个完全托管的环境，用于构建、训练和部署机器学习模型，包括自动模型调优 功能。具体而言，SageMaker包含一项名为自动模型调优（或超参数优化）的功能，它能够自动完成为您 的机器学习模型寻找最佳超参数的过程。当您希望在无需人工干预的情况下优化模型性能时，这一功能至 关重要。 亚马逊SageMaker能够让您自动化训练过程和超参数调整，这与该公司对完全自动化模型调优的需求 完美契合。

Human-in-the-loop (HITL) is a technique where human oversight is involved in the decision-making process of Al systems. In the context of generative Al applications, HITL can be used during the post- processing phase to identify and mitigate biases or toxic outputs. Humans can review and intervene when the model generates inappropriate or biased content, providing corrections or adjustments that help reduce the likelihood of toxicity and bias. This feedback loop helps refine and improve the model's outputs over time.

“人在回路”（HITL）是一种在人工智能系统的决策过程中引入人工监督的技术。在生成式人工智能应用 的背景下，HITL可用于后期处理阶段，以识别和减轻偏见或有害输出。当模型生成不适当或有偏见的内容 时，人类可以进行审查和干预，提供修正或调整，从而有助于降低有害性和偏见的可能性。这种反馈循环 有助于随着时间的推移不断改进模型的输出。

The issue described is a case of bias in the model's decision-making process, where the model is showing a preference for a specific demographic. The most cost-effective way to address this is to include more diverse training data that better represents all demographics. By fine-tuning the model with this more diverse data, you can help ensure that the model treats all demographic groups fairly and does not exhibit biased behavior. Fine-tuning the model with updated data ensures that the model learns from a more representative sample, improving fairness in its predictions without needing to completely retrain the model from scratch.

所描述的问题是模型决策过程中存在偏见，即模型对特定人群表现出偏好。解决这一问题最经济有效的方法 是纳入更多样化的训练数据，使其能更好地代表所有人群。通过用这些更多样化的数据对模型进行微调，您 可以帮助确保模型公平对待所有人群，不会表现出偏见行为。 使用更新的数据对模型进行微调，可确保模型从更具代表性的样本中学习，从而提高其预测的公平性，而 无需从头开始完全重新训练模型。

The company's chatbot needs predictions from the LLM to understand users' intent with minimal latency-Real-timeinference This scenario requires Real-time inference, as chatbots typically need immediate responses with low latency to provide an interactive user experience. A data processing job needs to query the LLM to process gigabytes of text files on weekends - Batchtransform This use case is best suited for Batch transform. Since the data processing job involves handling large amounts of data, and it is scheduled for weekends, batch processing can handle these large volumes efficiently. The company's engineering team needs to create an APl that can process small pieces of text content and provide low-latency predictions -Real-time inference Real-time inference is the correct choice here, as the APl needs to process text quickly and provide immediate responses, making real-time inference appropriate.

该公司的聊天机器人需要从大型语言模型（LLM）获取预测结果，以便以最小的延迟理解用 产的意图一一实时推理。主口 这种场景需要实时推理，因为聊天机器人通常需要以低延迟提供即时响应，以实现交互式的用户体验。 一项数据处理任务需要在周末查询大型语言模型来处理数十亿字节的文本文件一一批量转换 此用例最适合批处理转换。由于数据处理作业涉及大量数据处理，且安排在周未进行，批处理能够高效 地处理这些大量数据。 该公司的工程团队需要创建一个能够处理小段文本内容并提供低延迟预测的API一一实时推 理。 实时推理是此处的正确选择，因为该API需要快速处理文本并提供即时响应，这使得实时推理成为合 适之选。）

AWs CloudTrail is the AWS service designed for logging and monitoring API calls made to AWS services, including Amazon Bedrock. CloudTrail records detailed information about the API requests, including the identity of the requester, the time of the request, and the source IP address. This service is ideal for logging all requests made to the Amazon Bedrock APl and meets the logging requirement. Amazon s3 Intelligent-Tiering is a storage class designed for storing data that has unpredictable access patterns. It automatically moves data between two access tiers (frequent and infrequent) based on usage, which helps reduce costs while ensuring data is still available when needed. For retaining logs securely over 5 years at the lowest possible cost, this storage class provides an efficient way to handle long-term storage requirements without incurring unnecessary costs.

AWSCloudTrail是AWS提供的一项服务，用于记录和监控对AWS服务（包括AmazonBedrock）的API 调用。CloudTrail会详细记录API请求的相关信息，包括请求者的身份、请求时间以及源IP地址。该服务非 常适合记录对AmazonBedrockAPI的所有请求，并满足日志记录要求。 亚马逊S3智能分层存储类别专为存储访问模式难以预测的数据而设计。它会根据使用情况自动在两个访问 层级（频繁访问和不频繁访问）之间移动数据，这有助于降低成本，同时确保在需要时数据仍可访问。对 于以最低成本安全保留5年以上的日志而言，这种存储类别提供了一种高效的方式来满足长期存储需求， 同时避免不必要的成本支出。

Amazon Personalize is an AWS service specifically designed to build and deploy personalized recommendations for users. It allows you to create custom machine learning models for personalized user experiences, such as product recommendations, search results, or content suggestions. For an ecommerce company aiming to improve search engine recommendations tailored to each user, Amazon Personalize is the ideal choice as it leverages user behavior data and machine learning to deliver highly relevantrecommendations.

亚马逊个性化服务是亚马逊网络服务（AWS）推出的一项专门用于为用户构建和部署个性化推荐的服务。 它允许您创建定制的机器学习模型，以实现个性化的用户体验，例如产品推荐、搜索结果或内容建议。对 于一家希望改进针对每位用户量身定制的搜索引擎推荐的电子商务公司而言，亚马逊个性化服务是理想之 选，因为它利用用户行为数据和机器学习来提供高度相关的推荐。

Data residency refers to the practice of ensuring that data is stored and processed within a specific geographical location to comply with regulations or policies. In this scenario, where sensitive patient data must not leave the country it originates from, implementing a data residency strategy ensures compliance with legal and regulatory requirements while protecting patient privacy. This approach is crucial for organizations like hospitals operating under strict data governance frameworks such as HIPAA or GDPR.

数据驻留是指确保数据存储和处理在特定地理位置的一种做法，以符合法规或政策。在这种情况下，敏感 的患者数据不得离开其来源国，实施数据驻留策略可确保符合法律和监管要求，同时保护患者隐私。对于 像医院这样在严格的诸如《健康保险流通与责任法案》（HIPAA）或《通用数据保护条例》（GDPR）等 数据治理框架下运营的组织而言，这种方法至关重要。

Amazon CloudWatch is a highly scalable AWS service designed for monitoring and observability. It provides real-time monitoring of system metrics, including performance data for ML systems, such as resource utilization (CPU, memory, etc.), model inference latency, request counts, and errors. CloudWatch enables users to set up alarms, visualize metrics, and automate actions based on performance thresholds, making it ideal for monitoring the performance of ML systems.

亚马逊CloudWatch是一项高度可扩展的AWS服务，专为监控和可观测性而设计。它提供系统指标的实时 监控，包括机器学习系统的性能数据，例如资源利用率（CPU、内存等）、模型推理延迟、请求数量和错误。 CloudWatch允许用户设置警报、可视化指标，并根据性能值自动执行操作，因此非常适合用于监控机器 学习系统的性能。

Chain-of-thought prompting is a technique in prompt engineering where the Al model is encouraged to break down its reasoning process step by step to solve a problem, such as numerical reasoning challenges. By explicitly instructing the model to "show its work by explaining its reasoning step by step," the practitioner ensures the model provides a logical sequence of intermediate steps leading to the solution. This improves the accuracy and transparency of the model's outputs, particularly for complex reasoning tasks.

链式思维提示是提示工程中的一种技术，通过这种技术，人工智能模型被鼓励逐步分解其推理过程以解决 问题，例如数值推理挑战。通过明确指示模型“通过逐步解释其推理过程来展示其工作”，从业者可以确 保模型提供一系列合乎逻辑的中间步骤，从而得出解决方案。这提高了模型输出的准确性和透明度，特别 是在复杂的推理任务中。

Amazon Bedrock is an AwS service that allows users to build and scale generative Al applications using foundation models (FMs) without the need to manage infrastructure. It provides access to pre-trained models from AWS and third-party providers, enabling developers to integrate generative Al capabilities like text generation, summarization, and question-answering into their applications. This makes Amazon Bedrock the ideal choice for building and scaling generative Al solutions.

亚马逊Bedrock是一项AWS服务，它让用户能够使用基础模型（FMs）构建和扩展生成式人工智能应用程 序，而无需管理基础设施。它提供了来自AWS和第三方供应商的预训练模型的访问权限，使开发人员能够 将文本生成、摘要和问答等生成式人工智能功能集成到他们的应用程序中。这使得业马逊Bedrock成为构建 和扩展生成式人工智能解决方案的理想选择。

To meet the requirements of enabling the app to "hear what users say and provide voice responses," the solution must include speech recognition and text-to-speech capabilities. Using a deep learning neural network for speech recognition allows the app to convert spoken words into text. Once the input is understood, text-to-speech systems can provide voice responses back to users. This approach is fundamental in applications that assist users with visual impairments by enabling interaction through spoken language.

为了满足让应用程序能够“听懂用户所说的话并提供语音回复”的要求，解决方案必须包含语音识别和文 本转语音的功能。使用深度学习神经网络进行语音识别，应用程序能够将语音转换为文本。一旦理解了输 入内容，文本转语音系统就能向用户提供语音回复。这种方法对于帮助视障用户的应用程序来说至关重要， 因为它能让用户通过语音进行交互。

Chain-of-thought prompting is specifically designed to enhance the reasoning capabilities of a large language model (LLM) by encouraging it to provide step-by-step explanations for complex problem-solving tasks. This approach helps the model break down a problem into smaller, logical steps, ensuring that the response is detailed, accurate, and easy to follow.

链式思维提示专门用于增强大型语言模型（LLM）的推理能力，通过鼓励其为复杂的问题解决任务提供逐 步解释来实现。这种方法有助于模型将问题分解为更小的、合乎逻辑的步骤，确保响应详细、准确且易于理 解。

Continuous pre-training keeps a foundation model "always learning" by periodically (or even continuously) ingesting the latest data and updating its weights. This approach ensures the model stays current without having to start training from scratch every time new data arrives, unlike static or strictly batch-oriented methods.

持续的预训练通过定期（甚至持续）获取最新数据并更新其权重，使基础模型“始终处于学习状态”。这 种方法确保模型始终处于最新状态，而无需每次有新数据到来时都从头开始训练，这与静态或严格批量处 理的方法不同。

Binary classification: Supervised learning Binary classification involves predicting one of two possible outcomes, requiring labeled training data. Multi-classclassification:Supervised learning Multi-class classification extends binary classification to predict one of several classes, also requiring labeled data. K-means clustering: Unsupervised learning K-means clustering is used to group data points into clusters without requiring labeled data. Dimensionality reduction:Unsupervised learning Dimensionality reduction techniques, like PCA, reduce the number of features in the dataset without needing labeled data.

二分类：有监督学习 二分类涉及预测两种可能结果中的一种，需要有标签的训练数据。 多类别分类：有监督学习 多类别分类将二分类扩展到预测多个类别中的一个，同样需要有标签的数据。 K均值聚类：无监督学习 K均值聚类用于将数据点分组为簇，无需标注数据。 降维：无监督学习 诸如主成分分析（PCA）之类的降维技术可以在无需标注数据的情况下减少数据集中的特征数量。

Al governance frameworks focus on building trust and ensuring the responsible deployment of Al technologies. They establish clear policies and guidelines to address critical aspects such as data management, transparency, ethical considerations, responsible Al practices, and regulatory compliance. These frameworks help organizations mitigate risks, promote fairness, and foster public trust in Al systems, making them essential for creating human-centered Al technologies.

人工智能治理框架侧重于建立信任并确保人工智能技术的负责任部署。它们制定明确的政策和指导方 针，以解决诸如数据管理、透明度、伦理考量、负责任的人工智能实践以及监管合规等关键方面的问 题。这些框架有助于组织降低风险、促进公平，并增强公众对人工智能系统的信任，对于打造以人为 本的人工智能技术至关重要。

To measure the financial effect of a generative Al chatbot on the company's operations, the cost for each customer conversation is the most relevant metric. It directly quantifies the operational expense associated with using the chatbot to handle customer inquiries. By analyzing this metric, the company can evaluate how much it spends per conversation and compare it to the cost of alternative methods (e.g., human agents), providing insight into the chatbot's financial efficiency and ROl.

要衡量生成式人工智能聊天机器人对公司运营的财务影响，每个客户对话的成本是最相关的指标。它直接 量化了使用聊天机器人处理客户咨询所产生的运营费用。通过分析这一指标，公司可以评估每次对话的花 费，并将其与替代方法（例如人工客服）的成本进行比较，从而了解聊天机器人的财务效率和投资回报率。

K-means is a clustering algorithm used in unsupervised learning to group data points into clusters based on their similarities. It is well-suited for finding patterns in customer demographics and buying behavior. The algorithm identifies groups (clusters) of customers with similar characteristics, which can then be used for targeted marketing, personalized recommendations, or segmentation.

K均值算法是一种用于无监督学习的聚类算法，它根据数据点的相似性将数据点分组为不同的簇。该算法非 常适合用于发现客户人口统计信息和购买行为中的模式。通过该算法可以识别出具有相似特征的客户群体 （簇），然后可用于定向营销、个性化推荐或市场细分。

In the context of large language models (LLMs), the temperature inference parameter controls the randomness of the model's output. Lowering the temperature reduces randomness and makes the model's responses more deterministic and focused on the most likely predictions. By decreasing the temperature, the likelihood of hallucinations (when the model generates incorrect or nonsensical information) is reduced, as the model relies more on high-probability outputs rather than exploring less likely possibilities.

在大型语言模型（LLM）的语境中，温度推理参数控制着模型输出的随机性。降低温度会减少随机性，使模 型的输出更加确定。20 通过降低温度，响应会更具确定性，更侧重于最有可能的预测。温度降低后，模型产生错误或无意义信 息（即出现幻觉）的可能性会降低，因为模型更多地依赖高概率输出，而非探索不太可能的情况。

To give a large language model (LLM) the ability to use content from previous customer messages in a conversation, the previous messages must be included in the model prompt. This technique is known as prompt engineering and allows the LLM to retain context by incorporating a history of the interaction within the prompt. By appending prior exchanges to the prompt, the model can generate contextually relevant and coherent responses throughout the multi-turn conversation.

要让大型语言模型（LLM）能够使用先前客户消息中的内容进行对话，必须将之前的对话内容包含在 模型提示中。这种技术被称为提示工程，它通过在提示中加入交互历史，使LLM能够保留上下文。 通过将之前的交流内容附加到提示中，模型能够在多轮对话中生成与上下文相关且连贯的回复。

Amazon Bedrock enables companies to use foundation models (FMs) to build and automate tasks like generating product descriptions and recommendations. It allows the integration of pre-trained FMs into applications without managing infrastructure, making it an ideal choice for automating customer service tasks. With Amazon Bedrock, the company can leverage FMs to generate tailored recommendations based on customer locations, enabling dynamic and efficient customer interactions.

亚马逊Bedrock使企业能够利用基础模型（FMs）来构建和自动化诸如生成产品描述和推荐之类的任 务。它允许将预训练的基础模型集成到应用程序中，而无需管理基础设施，因此是自动执行客户服务 任务的理想选择。借助亚马逊Bedrock，企业可以利用基础模型根据客户所在位置生成定制推荐，从 而实现动态且高效的客户互动。

Amazon Macie is a fully managed data security and privacy service that uses machine learning to discover and protect sensitive data in Amazon S3. It can automatically detect sensitive information, such as personally identifiable information (Pll) or financial data, and send alerts when such data is found. This approach minimizes development effort, as it does not require custom regex patterns or model development, and it is specifically designed to handle the scenario described.

AmazonMacie是一项完全托管的数据安全和隐私服务，它利用机器学习在AmazonS3中发现并保护敏 感数据。它可以自动检测诸如个人身份信息（PI）或财务数据之类的敏感信息，并在发现此类数据时 发送警报。这种方法最大限度地减少了开发工作量，因为它无需自定义正则表达式模式或模型开发， 并且专门针对所描述的场景进行了设计。

"claeeifrthnfallaiainn tavt ae aithnranartenalitinear nntartainmnnt,rinniittavti " MM CorrectAnswer:Zero-shot learning Zero-shot learning involves providing the model with a task and no prior examples, relying entirely on the model's pre-trained knowledge to perform the classification. "A [image 1], [image 2], and [image 3] are examples of [target class]. Classify the following image nAT+APNAAIAAAT" Correct Answer: Few-shot learning Few-shot learning provides the model with a few examples (image 1, image 2, and image 3) before asking it to classify a new instance, helping it generalize to the task with minimal examples. "riiaetian1rinetriirtiane tafallaaiThinketanhuietan anduialk mnthranahanrthinina" Ln Irtr rT sToJao Correct Answer: Chain-of-thought reasoning Chain-of-thought reasoning encourages the model to break down its reasoning process step by step, enhancing its ability to solve complex tasks logically.

“将以下文本分类为体育、政治或娱乐：【输入文本]。”正确答案：零样本学习 零样本学习是指仅向模型提供任务而无任何先前示例，完全依赖模型的预训练知识来进行分类。 “[图片1]、[图片2] 和[图片3] 是[目标类别] 的示例。将以下图片分类为[目标类别]。” 少样本学习 少样本学习在要求模型对新实例进行分类之前，先为其提供少量示例（图像1、图像2和图像3）， 帮助其仅通过少量宗例就能泛花到该在务。的，一一一的 [问题。I[遵循的说明。1逐步思考，并向我阐述你的思路。正确答案：链式思维推理 链式思维推理鼓励模型逐步分解其推理过程，从而增强其逻辑解决复杂任务的能力。

Block input prompts or model responses that contain harmful content such as hate, insults, violence, or misconduct - Content filters Avoid subjects related to illegal investment advice or legal advice- Denied topics Detect and block specific offensive terms-Word filters Detect and filter out information in the model's responses that is not grounded in the provided source information- Contextual grounding check

屏蔽包含仇恨、侮辱、暴力或不当行为等有害内容的输入提示或模型响应-内容过滤器 避免涉及非法投资建议或法律咨询相关的话题－禁止主题 检测并屏蔽特定的冒犯性词汇-词库过滤器 检测并过滤掉模型回复中未基于所提供内容的信息。 源信息-上下文关联检查

Amazon SageMaker Model Cards provide a structured way to document key details about Al models, including their intended use, performance metrics, and limitations. This helps organizations maintain transparency, compliance, and governance in Al model development, making it easier to track and manage models over time.

亚马逊SageMaker模型卡片提供了一种结构化的方式来记录有关AI模型的关键细节，包括其预期用途、 性能指标和局限性。这有助于组织在AI模型开发过程中保持透明度、合规性和治理能力，从而更轻松 地跟踪和管理模型。

The F1 score is a metric used to evaluate the performance of a classification model by considering both precision (the proportion of correctly predicted positive cases out of all predicted positives) and recall (the proportion of correctly predicted positive cases out of all actual positives). It is the harmonic mean of precision and recall, ensuring a balance between them, especially when dealing with imbalanced datasets.

F1分数是一种用于评估分类模型性能的指标，它同时考虑了准确率（所有预测为正例的样本中实际为正例 的比例）和召回率（所有实际为正例的样本中被正确预测为正例的比例）。它是准确率和召回率的调和平均 值，确保了两者之间的平衡，尤其是在处理不平衡数据集时。

Retrieval Augmented Generation (RAG) is a cost-effective approach that enhances Al-generated responses by retrieving relevant information from external knowledge sources. Instead of fine-tuning or re- training a model, RAG dynamically pulls the most recent and relevant data at query time. This is particularly useful in scenarios where questions change over time, ensuring that the Al/ML solution provides accurate and up-to-date responses without requiring expensive and time-consuming model retraining. Prompt engineering techniques further optimize how the model processes and generates responses, improving accuracy and relevance.

检索增强生成（RAG）是一-种经济高效的方法，通过从外部知识源检索相关信息来增强AI生成的响应。与 对模型进行微调或重新训练不同，RAG在查询时动态获取最新且最相关的数据。这在问题随时间变化的场 景中特别有用，确保AI/ML解决方案能够提供准确且最新的响应，而无需进行昂贵且耗时的模型重新训练。 提示工程技术进一步优化了模型处理和生成响应的方式，提高了准确性和相关性。

The scenario describes a dataset that is not representative of all demographics, which can lead to biased model predictions. This directly relates to fairness, a core dimension of responsible Al that ensures Al systems make unbiased and equitable decisions across different demographic groups. Addressing fairness involves techniques such as balanced dataset curation, bias detection, and mitigation strategies to ensure that the Al system does not discriminate against any group.

该场景描述了一个不能代表所有人口群体的数据集，这可能导致模型预测出现偏差。这直接关系到公平性，这是负责任的人工智能 的一个核心维度，旨在确保人工智能系统在不同人口群体中做出无偏见且公平的决策。解决公平性问题涉及诸如平衡数据集的整理、 偏差检测以及缓解策略等技术手段，以确保人工智能系统不会对任何群体产生歧视。

Aws Audit Manager helps organizations continuously assess and audit compliance with industry regulations and internal policies by automating evidence collection and generating audit reports. This is essential for ensuring that the ML application meets regulatory requirements. Aws Config enables continuous monitoring and compliance checks by tracking configuration changes in Aws resources. It helps the company ensure that infrastructure settings align with security policies and industry standards.

解释; AWS审计管理器通过自动化证据收集和生成审计报告，帮助组织持续评估和审计其对行业法规和内部政策的合规性。这对于： 确保机器学习应用程序符合监管要求至关重要。 AWSConfig通过跟踪AWS资源的配置更改，实现持续监控和合规性检查。它有助于公司确保基础设施设置符合安全策略和 行业标准。

Using prompt engineering is the most cost-effective way to improve the accuracy of responses from a generative Al application without retraining or fine-tuning the foundation model (FM). Prompt engineering involves carefully designing the input prompts to guide the model toward producing better responses, improving relevance and accuracy.

使用提示工程是在不重新训练或微调基础模型（FM）的情况下提高生成式人工智能应用程序响应准确性的最具成本效益的方法。 提示工程涉及精心设计输入提示，以引导模型生成更优质的响应。 提高相关性和准确性。