AWS AIP-C01 开源练习 · 基础 · 第 01 套

A. The Generative AI Lens extends the Well-Architected Framework with design principles and best practices tailored to GenAI workloads.

B. The core pillars are general-purpose; the GenAI-specific guidance lives in the Generative AI Lens.

C. Trusted Advisor gives cost/security/limits checks, not GenAI architectural best practices.

D. Config evaluates resource compliance against rules; it is not a source of GenAI design guidance.

A. 生成式 AI 镜头通过针对 GenAI 工作负载量身定制的设计原则和最佳实践扩展了架构完善的框架。

B. 核心支柱是通用的； GenAI 特定的指导存在于 Generative AI Lens 中。

C. Trusted Advisor 提供成本/安全/限制检查，而不是 GenAI 架构最佳实践。

D. Config 根据规则评估资源合规性；它不是 GenAI 设计指导的来源。

A. The WA Tool runs a structured review against the framework and any selected lens, surfaces HRIs, and generates an improvement plan.

B. CloudWatch monitors runtime metrics; it does not perform architectural reviews or produce improvement plans.

C. Cost Explorer analyzes spend, not architectural risk.

D. Artifact provides compliance reports/agreements, not workload reviews.

A. WA 工具对框架和任何选定的镜头进行结构化审查，显示 HRI，并生成改进计划。

B. CloudWatch监控运行时指标；它不执行架构审查或制定改进计划。

C. Cost Explorer 分析支出，而不是架构风险。

D. Artifact 提供合规性报告/协议，而不是工作负载审查。

A. Graceful degradation, failover, and handling throttling are Reliability-pillar concerns.

B. Cost Optimization is about spending efficiently, not resilience to load.

C. Sustainability targets environmental impact, not failure handling.

D. Operational Excellence covers run/observe/improve processes; resilient behavior under failure maps to Reliability.

A. 优雅降级、故障转移和处理限制是可靠性支柱问题。

B. 成本优化是关于有效支出，而不是负载弹性。

C. 可持续性的目标是环境影响，而不是故障处理。

D. 卓越运营涵盖运行/观察/改进流程；故障下的弹性行为映射到可靠性。

A. Right-sizing model choice, prompt routing, and distillation to cut inference cost are Cost Optimization practices.

B. Security covers protecting data and workloads, not spend.

C. Reliability is about resilience and recovery, not cost.

D. Performance Efficiency is related (model selection) but the explicit goal here is reducing cost.

A. 正确调整模型大小、快速路由和精炼以降低推理成本是成本优化实践。

B. 安全性涵盖保护数据和工作负载，而不是保护支出。

C. 可靠性涉及弹性和恢复，而不是成本。

D. 性能效率是相关的（模型选择），但这里的明确目标是降低成本。

A. Circuit breakers prevent cascading failures by halting calls to a failing dependency.

B. Cross-Region failover keeps the workload available during a regional impairment.

C. Temperature affects output randomness, not reliability.

D. A single-Region store is a single point of failure, the opposite of resilient design.

E. Removing retries reduces, not improves, resilience to transient failures.

A. 断路器通过停止对失败依赖项的调用来防止级联故障。

B. 跨区域故障转移可在区域故障期间保持工作负载可用。

C. 温度影响输出随机性，而不是可靠性。

D. 单区域存储是单点故障，与弹性设计相反。

E. 取消重试会降低而不是提高对瞬时故障的恢复能力。

A. Falling back to a reduced-quality but functional response when a dependency fails is graceful degradation.

B. Provisioned throughput reserves model capacity; it is not a fallback pattern.

C. Blue/green is a deployment-cutover strategy, unrelated to runtime fallback.

D. That is a model-training technique, not a resilience pattern.

A. 当依赖失败时，回退到质量降低但功能正常的响应是优雅的降级。

B. 预置吞吐量预留模型容量；这不是后备模式。

C. 蓝/绿是一种部署切换策略，与运行时回退无关。

D. 这是一种模型训练技术，而不是弹性模式。

A. Data protection, least-privilege IAM, and content safeguards (guardrails) are Security-pillar best practices.

B. Operational Excellence is about running and improving workloads, not data protection controls.

C. Performance Efficiency targets effective resource use, not protection.

D. Reliability concerns availability and recovery, not access control and encryption.

A. 数据保护、最低权限 IAM 和内容保护（护栏）是安全支柱最佳实践。

B. 卓越运营是指运行和改进工作负载，而不是数据保护控制。

C. 性能效率的目标是有效地利用资源，而不是保护。

D. 可靠性涉及可用性和恢复，而不是访问控制和加密。

A. Versioned model catalogs feeding CI/CD are an Operational Excellence practice; the Model Registry provides this.

B. Budgets track spend; they do not provide model versioning or deployment automation.

C. WAF protects web apps from exploits; it is unrelated to model versioning.

D. Route 53 is DNS/failover; it does not manage model versions or CI/CD.

A. 提供 CI/CD 的版本化模型目录是一种卓越运营实践；模型注册表提供了这一点。

B. 预算跟踪支出；它们不提供模型版本控制或部署自动化。

C. WAF 保护 Web 应用程序免受攻击；它与模型版本控制无关。

D. 路由 53 是 DNS/故障转移；它不管理模型版本或 CI/CD。

A. AppConfig externalizes runtime configuration (such as the model ID) so it can change without a code deployment.

B. CodeDeploy automates deployments of new code/artifacts, which is exactly what the team wants to avoid here.

C. Provisioned Throughput reserves model capacity; it does not switch models via config.

D. Patch Manager handles OS patching, unrelated to model selection.

A. AppConfig 外部化运行时配置（例如模型 ID），因此无需部署代码即可更改它。

B. CodeDeploy 自动部署新代码/工件，这正是团队想要避免的。

C. 预配置吞吐量储备模型容量；它不通过配置切换模型。

D. 补丁管理器处理操作系统补丁，与模型选择无关。

A. Right-sizing models and maximizing utilization of managed infrastructure reduces energy/resource consumption.

B. Oversized models waste compute and energy, contradicting sustainability.

C. Keeping idle accelerators running wastes resources.

D. Unnecessary global duplication increases resource consumption.

A. 正确调整模型规模并最大限度地利用托管基础设施可以减少能源/资源消耗。

B. 过大的模型会浪费计算和能源，与可持续性相矛盾。

C. 保持闲置加速器运行会浪费资源。

D. 不必要的全局重复会增加资源消耗。

A. HRIs are the highest-priority findings the WA Tool surfaces for an improvement plan.

B. HRIs come from the architectural review, not billing analysis.

C. HRIs are review findings about design, not runtime model outputs.

D. HRIs are risk findings, not quota requests.

A. HRI 是 WA 工具为改进计划提供的最高优先级发现。

B. HRI 来自架构审查，而不是计费分析。

C. HRI 是关于设计的审查结果，而不是运行时模型输出。

D. HRI 是风险发现，而不是配额请求。

A. Selecting the right model and using caching to meet latency targets is core Performance Efficiency.

B. While related, the explicit goal here is meeting performance targets, not environmental impact.

C. Security is unrelated to latency optimization.

D. Reliability addresses availability/recovery, not latency tuning.

A. 选择正确的模型并使用缓存来满足延迟目标是核心性能效率。

B. 虽然相关，但这里的明确目标是满足绩效目标，而不是环境影响。

C. 安全性与延迟优化无关。

D. 可靠性涉及可用性/恢复，而不是延迟调整。

A. Guardrails enforce content policies, a key responsible-AI safeguard.

B. Removing PII before use limits exposure of sensitive data, a Security best practice.

C. Wildcard permissions violate least privilege.

D. Disabling audit logging undermines security and compliance.

E. Turning off encryption directly violates the Security pillar.

A. 护栏执行内容策略，这是负责任的人工智能的一项关键保障措施。

B. 在使用前删除 PII 可限制敏感数据的暴露，这是一种安全最佳实践。

C. 通配符权限违反了最小权限。

D. 禁用审核日志记录会破坏安全性和合规性。

E. 关闭加密直接违反了安全支柱。

A. Lenses layer specialized guidance (e.g., GenAI, Serverless, ML) on top of the six general pillars.

B. Lenses augment, not replace, the core pillars.

C. Lenses are guidance, not pricing mechanisms.

D. Lenses are review guidance, not enforced runtime controls.

A. 镜头在六个一般支柱之上提供了专门的指导（例如 GenAI、Serverless、ML）。

B. 镜头增强而不是取代核心支柱。

C. 镜头是指导，而不是定价机制。

D. 镜头是审查指导，而不是强制的运行时控制。

A. Bedrock exposes one API surface (e.g., InvokeModel/Converse) across many providers' models.

B. Bedrock's value is one API, not per-provider SDKs.

C. Bedrock is a fully managed service; you never access provider servers.

D. Bedrock is serverless and managed; self-hosting is not required.

A. Bedrock 在许多提供商的模型中公开一个 API 表面（例如 InvokeModel/Converse）。

B. Bedrock 的价值在于一个 API，而不是每个提供商的 SDK。

C. Bedrock 是一项完全托管的服务；您永远不会访问提供商服务器。

D. Bedrock 是无服务器且托管的；不需要自托管。

A. Bedrock model access is a per-Region setting; enabling it in one Region does not enable it elsewhere.

B. Models are available in many Regions; access is just granted per Region.

C. Partitions are unrelated; this is a per-Region access grant.

D. On-demand access works once granted; the issue is the per-Region grant, not throughput.

A. 基岩模型访问是按区域设置的；在一个区域启用它不会在其他地方启用它。

B. 许多地区都有型号可供选择；仅按区域授予访问权限。

C. 分区不相关；这是按区域的访问授权。

D. 一旦获得授权，即可按需访问；问题是每个区域的授权，而不是吞吐量。

A. Cross-Region inference profiles automatically route invocations across Regions to absorb spikes, with no code or model change.

B. Prompt routing uses different models in a family; the requirement is the same model.

C. Provisioned Throughput suits steady high volume and adds cost for occasional peaks.

D. A custom failover Lambda adds operational overhead vs the built-in profile.

A. 跨区域推理配置文件会自动跨区域路由调用以吸收峰值，无需更改代码或模型。

B. 快速路由使用一个系列中的不同型号；要求是同型号。

C. 预置吞吐量适合稳定的高容量，但会增加偶尔高峰的成本。

D. 与内置配置文件相比，自定义故障转移 Lambda 增加了运营开销。

A. Application inference profiles are custom profiles you create and tag, enabling cost tracking; system-defined profiles are AWS-managed.

B. System-defined profiles are predefined by AWS and are not user-tagged; application profiles are the taggable, custom ones.

C. That detects spend anomalies; it does not create taggable inference routing profiles.

D. Model Units reserve capacity; they are not the per-team tagging mechanism for cross-Region inference.

A. 应用程序推理配置文件是您创建和标记的自定义配置文件，可实现成本跟踪；系统定义的配置文件由 AWS 管理。

B. 系统定义的配置文件由 AWS 预定义，没有用户标记；应用程序配置文件是可标记的、自定义的。

C. 检测支出异常；它不会创建可标记的推理路由配置文件。

D. 模型单位储备容量；它们不是用于跨区域推理的每团队标记机制。

A. This condition key scopes access to a particular inference profile ARN.

B. That enforces guardrail usage, not inference-profile scoping.

C. That scopes access to a prompt router, not an inference profile.

D. That restricts by VPC endpoint, unrelated to inference profiles.

A. 此条件键范围限制对特定推理配置文件 ARN 的访问。

B. 这强制使用护栏，而不是推理配置文件范围。

C. 这限制了对提示路由器的访问，而不是推理配置文件。

D. 这受 VPC 端点限制，与推理配置文件无关。

A. Provisioned Throughput guarantees dedicated capacity (per Model Unit) for steady, predictable workloads.

B. On-demand is convenient but does not guarantee reserved capacity at peak.

C. Prompt routing optimizes cost/quality across a family; it does not reserve capacity.

D. Profiles spread load but do not reserve guaranteed capacity per model unit.

A. 预置吞吐量保证了稳定、可预测的工作负载的专用容量（每个模型单元）。

B. 按需提供虽然方便，但不能保证高峰时的预留容量。

C. 快速路由可优化整个家庭的成本/质量；它不保留容量。

D. 配置文件分散负载，但不保留每个模型单元的保证容量。

A. Customized (fine-tuned) Bedrock models are served via Provisioned Throughput, which is required to host them.

B. Custom models require Provisioned Throughput to serve, unlike base models on-demand.

C. Bedrock manages serving; manual EC2 hosting is not the mechanism.

D. Profiles route base-model traffic; they are not the requirement for serving a custom model.

A. 定制（微调）基岩模型通过托管它们所需的预配置吞吐量提供服务。

B. 与按需基本模型不同，自定义模型需要预置吞吐量才能提供服务。

C. Bedrock 管理服务；手动 EC2 托管不是该机制。

D. 配置文件路由基本模型流量；它们不是服务自定义模型的要求。

A. A prompt router analyzes each prompt and routes to the best model within the same family via a single serverless endpoint, reducing cost.

B. That reserves capacity but does not route by complexity or save cost on easy prompts.

C. Hand-rolled routing adds maintenance; the managed prompt router is the intended solution.

D. Profiles spread load across Regions; they do not route by prompt complexity.

A. 提示路由器分析每个提示并通过单个无服务器端点路由到同一系列中的最佳模型，从而降低成本。

B. 这保留了容量，但不会按复杂性进行布线，也不会根据简单的提示节省成本。

C. 手工布线增加了维护；托管提示路由器是预期的解决方案。

D. 配置文件将负载分散到各个区域；它们不会按即时复杂性进行路由。

A. This condition key scopes invocation to a specific prompt router ARN.

B. That scopes inference profiles, not prompt routers.

C. There is no router-scoping behavior in a plain model-id key for this purpose.

D. That pins a guardrail version, unrelated to routers.

A. 此条件键的范围是对特定提示路由器 ARN 的调用。

B. 该范围适用于推理配置文件，而不是提示路由器。

C. 为此目的，普通 model-id 键中没有路由器作用域行为。

D. 这固定了护栏版本，与路由器无关。

A. Prompt routers operate within a single family (e.g., Anthropic, Llama, or Nova variants).

B. Routing is constrained to one model family, not arbitrary cross-provider mixing.

C. Routing is serverless and does not mandate provisioned capacity.

D. Prompt routing targets text-generation models, not embeddings.

A. 提示路由器在单个系列中运行（例如，Anthropic、Llama 或 Nova 变体）。

B. 路由仅限于一个模型系列，而不是任意的跨提供商混合。

C. 路由是无服务器的，并且不强制要求配置容量。

D. 提示路由的目标是文本生成模型，而不是嵌入。

A. Fine-tuning trains on labeled input/output pairs to specialize a model for a task.

B. Continued pre-training uses unlabeled domain text, not labeled task examples.

C. Distillation transfers knowledge from a teacher to a smaller student; it is not the labeled-data adaptation described.

D. RAG augments prompts at inference time; it does not train the model.

A. 对标记的输入/输出对进行微调训练，以专门针对任务的模型。

B. 继续预训练使用未标记的域文本，而不是标记的任务示例。

C. 蒸馏将知识从教师转移给较小的学生；它不是所描述的标记数据适应。

D. RAG 增强了推理时的提示；它不训练模型。

A. Continued pre-training ingests unlabeled domain data to build domain familiarity.

B. Fine-tuning needs labeled task pairs; here the data is unlabeled.

C. Prompt engineering does not add domain knowledge to the model weights.

D. Distillation compresses a model; it does not inject a new unlabeled domain corpus.

A. 持续的预训练会摄取未标记的领域数据以建立领域熟悉度。

B. 微调需要标记任务对；这里的数据是未标记的。

C. 即时工程不会将领域知识添加到模型权重中。

D. 蒸馏压缩模型；它不会注入新的未标记域语料库。

A. Distillation transfers a teacher model's behavior into a smaller student for lower cost/latency.

B. That broadens domain knowledge but does not compress a teacher into a student.

C. That reserves capacity; it does not produce a smaller model.

D. Hybrid search is a retrieval technique, not a model-compression method.

A. 蒸馏将教师模型的行为转移到较小的学生中，以降低成本/延迟。

B. 这拓宽了领域知识，但不会将教师压缩为学生。

C. 即储备能力；它不会生产更小的模型。

D. 混合搜索是一种检索技术，而不是模型压缩方法。

A. LoRA injects low-rank matrices so only a small set of parameters is trained, cutting cost.

B. Adapter modules add a few trainable layers while the base model stays frozen, enabling multiple variants.

C. Full retraining is the opposite of parameter-efficient; it updates all weights at high cost.

D. Temperature is an inference setting, not a training technique.

E. Inference mode is unrelated to PEFT training.

A. LoRA 注入低秩矩阵，因此仅训练一小部分参数，从而降低成本。

B. 适配器模块添加了一些可训练层，同时基本模型保持冻结，从而实现多种变体。

C. 完全再训练与参数高效相反；它以高成本更新所有权重。

D. 温度是一种推理设置，而不是一种训练技术。

E. 推理模式与 PEFT 训练无关。

A. PEFT lets you keep one base model and swap small adapters, reducing training cost and storage per variant.

B. No training method guarantees zero hallucination.

C. PEFT still requires task data; it just trains fewer parameters.

D. Serving customized models still uses Provisioned Throughput.

A. PEFT 可让您保留一个基本模型并更换小型适配器，从而降低每个变体的培训成本和存储成本。

B. 没有任何训练方法可以保证零幻觉。

C. PEFT仍然需要任务数据；它只是训练更少的参数。

D. 为定制模型提供服务仍使用预配置吞吐量。

A. The Model Registry stores versioned models with approval states and integrates with CI/CD pipelines.

B. Prompt Management versions prompts, not trained models.

C. CodeArtifact is a software package repository, not a model registry.

D. ECR stores container images; the model-versioning/approval catalog is the Model Registry.

A. 模型注册表存储具有批准状态的版本化模型，并与 CI/CD 管道集成。

B. 提示管理版本提示，而不是经过训练的模型。

C. CodeArtifact 是一个软件包存储库，而不是模型注册表。

D. ECR存储容器镜像；模型版本控制/批准目录是模型注册表。

A. Bedrock model evaluation runs automatic and human evaluations to compare models on quality and responsible-AI metrics.

B. That reserves capacity; it does not evaluate model quality.

C. Glue Data Quality checks datasets, not model output quality.

D. Profiles route traffic; they do not evaluate models.

A. Bedrock 模型评估运行自动评估和人工评估，以比较模型的质量和负责任的 AI 指标。

B. 即储备能力；它不评估模型质量。

C. Glue 数据质量检查数据集，而不是模型输出质量。

D. 配置文件路由流量；他们不评估模型。

A. On-demand bills per request and needs no capacity reservation, ideal for variable low/moderate usage.

B. Provisioned Throughput reserves (and bills for) capacity, wasteful for spiky low usage.

C. Self-managed GPUs add cost and operational overhead versus on-demand Bedrock.

D. Batch transform is for offline bulk jobs, not interactive on-demand calls.

A. 按请求按需计费，无需容量预留，非常适合可变的低/中等使用量。

B. 预配置吞吐量保留（和计费）容量，对于尖峰低使用率来说是浪费。

C. 与按需 Bedrock 相比，自管理 GPU 增加了成本和运营开销。

D. 批量转换适用于离线批量作业，而不是交互式按需调用。

A. Batch inference processes large datasets asynchronously at lower cost than real-time, ideal when latency does not matter.

B. Real-time per-call invocation is costlier and slower for huge offline volumes.

C. Reserving real-time capacity is unnecessary and costlier for an offline batch.

D. Routing optimizes per-prompt model choice; it is not a bulk offline mechanism.

A. 批量推理以比实时更低的成本异步处理大型数据集，是延迟不重要时的理想选择。

B. 对于巨大的离线量来说，实时的每次调用调用成本更高且速度更慢。

C. 对于离线批量来说，预留实时容量是不必要的，而且成本更高。

D. 路由优化了每个提示的模型选择；它不是批量离线机制。

A. Custom Model Import lets you bring supported open-weight custom models and serve them via Bedrock.

B. That trains a Bedrock base model further; it does not import external weights.

C. Prompt Flows orchestrate prompts; they do not import models.

D. Knowledge Bases provide RAG, not model import.

A. 自定义模型导入可让您引入受支持的开放重量自定义模型并通过 Bedrock 提供服务。

B. 进一步训练基岩基础模型；它不导入外部权重。

C. 提示流编排提示；他们不导入模型。

D. 知识库提供 RAG，而不是模型导入。

A. Right-sizing to requirements balances cost, latency, and quality rather than defaulting to the biggest model.

B. Bigger is not always better; it raises cost/latency without guaranteed task benefit.

C. Recency alone is not a selection criterion; fit to requirements matters.

D. Brand is not a substitute for evaluating fit, cost, and performance.

A. 根据需求调整规模可以平衡成本、延迟和质量，而不是默认使用最大的模型。

B. 越大并不总是越好；它会增加成本/延迟，而无法保证任务收益。

C. 仅仅新近度并不是选择标准；符合要求很重要。

D. 品牌不能替代评估适合性、成本和性能。

A. Titan and Nova are Amazon-built FMs accessible via the Bedrock unified API, including multimodal Nova variants.

B. Titan/Nova are Amazon's own models, not Anthropic's.

C. They are cloud models served through Bedrock.

D. They are foundation models, not IAM tooling.

A. Titan 和 Nova 是 Amazon 构建的 FM，可通过 Bedrock 统一 API 访问，包括多模式 Nova 变体。

B. Titan/Nova 是亚马逊自己的模型，而不是 Anthropic 的。

C. 它们是通过 Bedrock 提供服务的云模型。

D. 它们是基础模型，而不是 IAM 工具。

A. Custom/fine-tuned Bedrock models are hosted via Provisioned Throughput.

B. Custom models need Provisioned Throughput, unlike base models on-demand.

C. Bedrock supports fine-tuning for supported models.

D. Bedrock manages the training; EC2 migration is not required.

A. 自定义/微调的基岩模型通过预配置吞吐量托管。

B. 与按需基本模型不同，自定义模型需要预置吞吐量。

C. Bedrock 支持对支持的模型进行微调。

D. Bedrock 管理培训；不需要 EC2 迁移。

A. Circuit breakers protect the system from a failing dependency.

B. Graceful degradation preserves a usable experience during failures.

C. Unbounded waits worsen reliability and can exhaust resources.

D. A single Region with no failover is a single point of failure.

E. top-p is an inference parameter and has nothing to do with availability.

A. 断路器可以保护系统免受失败的依赖关系的影响。

B. 优雅的降级可以在故障期间保留可用的体验。

C. 无限制的等待会降低可靠性并耗尽资源。

D. 没有故障转移的单个区域是单点故障。

E. top-p是一个推理参数，与可用性无关。

A. System-defined = automatic/AWS-managed; application = custom and taggable for cost allocation.

B. This reverses the definitions.

C. They differ in who creates them and whether they support custom tagging.

D. Both can support cross-Region routing; the distinction is creation/tagging.

A. 系统定义=自动/AWS管理；应用程序 = 自定义且可标记以进行成本分配。

B. 这颠倒了定义。

C. 它们的不同之处在于谁创建了它们以及它们是否支持自定义标记。

D. 两者都可以支持跨Region路由；区别在于创建/标记。

A. Application inference profiles add taggable, attributable usage for cost tracking across applications.

B. That reserves capacity and changes cost model but does not provide per-app tagging by itself.

C. Removing logging reduces visibility, the opposite of what is needed.

D. Sharing identities harms attribution and security.

A. 应用程序推理配置文件添加了可标记、可归因的使用情况，用于跨应用程序的成本跟踪。

B. 这保留了容量并改变了成本模型，但本身不提供每个应用程序标记。

C. 删除日志记录会降低可见性，这与所需的相反。

D. 共享身份会损害归属和安全。

A. Provisioned Throughput shines for steady, high, predictable demand.

B. On-demand is cheaper for sporadic low usage.

C. Batch inference is better for one-off offline volume.

D. Cross-Region inference profiles better absorb occasional spikes without reserving capacity.

A. 预配置吞吐量因稳定、高且可预测的需求而大放异彩。

B. 对于零星的低使用量来说，按需服务更便宜。

C. 批量推理更适合一次性离线量。

D. 跨区域推理配置文件可以更好地吸收偶尔的峰值，而无需保留容量。

A. AppConfig externalizes the model choice for safe, no-redeploy, rollback-friendly changes.

B. Redeploying for each change is exactly what AppConfig avoids.

C. Layers still require redeployment to change values.

D. Provisioned Throughput reserves capacity; it is not a config switch.

A. AppConfig 将模型选择具体化，以实现安全、无需重新部署、易于回滚的更改。

B. AppConfig 恰恰避免了针对每次更改进行重新部署。

C. 层仍然需要重新部署才能更改值。

D. 预留吞吐量储备能力；它不是一个配置开关。

A. Distillation trades a little quality for big cost/latency savings by compressing the teacher into a student.

B. Distillation reduces size, not increases it.

C. That describes continued pre-training, not distillation.

D. No technique guarantees zero hallucination.

A. 蒸馏通过将教师压缩为学生，以牺牲一点质量来节省大量成本/延迟。

B. 蒸馏会减小尺寸，而不是增大尺寸。

C. 这描述的是持续的预训练，而不是蒸馏。

D. 没有任何技术可以保证零幻觉。

A. The condition key restricts invocation to the approved inference profile ARN.

B. Code-only controls are not enforceable security; use IAM conditions.

C. That enforces guardrails, not inference-profile scoping.

D. That is a blunt change unrelated to scoping a specific profile.

A. 条件键将调用限制为已批准的推理配置文件 ARN。

B. 仅代码控制不是可执行的安全性；使用 IAM 条件。

C. 这加强了护栏，而不是推理配置文件范围。

D. 这是一个与确定特定配置文件范围无关的直接更改。

A. Glue Data Quality computes rules and automatically detects anomalies in datasets within Glue pipelines.

B. Textract extracts text from documents; it does not measure dataset quality.

C. Transcribe converts speech to text; unrelated to data-quality checks.

D. Comprehend performs NLP analysis, not pipeline data-quality scoring.

A. Glue 数据质量计算规则并自动检测 Glue 管道内数据集中的异常情况。

B. Textract 从文档中提取文本；它不衡量数据集质量。

C. Transcribe 将语音转换为文本；与数据质量检查无关。

D. Comprehend 执行 NLP 分析，而不是管道数据质量评分。

A. Data Wrangler is the low-code GUI for importing, visualizing, and transforming data for ML.

B. Processing runs code-based processing jobs; it is not the GUI prep tool.

C. Glue Data Quality scores data quality; it is not a visual data-prep canvas.

D. Comprehend analyzes text; it is not a tabular data-prep GUI.

A. Data Wrangler 是用于导入、可视化和转换 ML 数据的低代码 GUI。

B. 处理运行基于代码的处理作业；它不是 GUI 准备工具。

C. Glue Data Quality 对数据质量进行评分；它不是一个可视化的数据准备画布。

D. 理解分析文本；它不是一个表格数据准备 GUI。

A. SageMaker Processing runs containerized data-processing/feature-engineering jobs on managed compute.

B. Data Wrangler is interactive GUI prep; Processing is for programmatic jobs at scale.

C. Transcribe is speech-to-text, not a processing-job service.

D. Kendra is enterprise search, not a data-processing job runner.

A. SageMaker Processing 在托管计算上运行容器化数据处理/特征工程作业。

B. Data Wrangler 是交互式 GUI 准备工具；处理适用于大规模的程序化作业。

C. 转录是语音到文本的服务，而不是处理作业服务。

D. Kendra 是企业搜索，而不是数据处理作业运行程序。

A. Transcribe converts audio speech to text, ideal for preparing call transcripts.

B. Polly is text-to-speech, the reverse operation.

C. Textract extracts text from documents/images, not audio.

D. Comprehend analyzes text; it does not transcribe audio.

A. 转录将音频语音转换为文本，非常适合准备通话记录。

B. Polly 是文本转语音，即相反的操作。

C. Textract 从文档/图像中提取文本，而不是音频。

D. 理解分析文本；它不转录音频。

A. Textract extracts printed/handwritten text, forms, and tables from scanned documents.

B. Transcribe handles audio, not documents.

C. Comprehend analyzes already-extracted text; it does not OCR documents.

D. Rekognition analyzes images/video for objects and faces, not structured document text extraction.

A. Textract 从扫描文档中提取打印/手写文本、表格和表格。

B. Transcribe 处理音频，而不是文档。

C. 理解分析已提取的文本；它不支持 OCR 文档。

D. Rekognition 分析图像/视频中的对象和面部，而不是结构化文档文本提取。

A. Comprehend detects and can redact PII entities from text, suited to pre-processing for FMs.

B. Macie discovers/classifies sensitive data in S3 for security posture; it is not a text PII-redaction step for GenAI pipelines.

C. Textract extracts text; it does not redact PII.

D. Transcribe converts audio to text; it is not a PII redactor for documents.

A. Comprehend 检测并可以从文本中编辑 PII 实体，适合 FM 的预处理。

B. Macie 发现/分类 S3 中的敏感数据以确保安全；它不是 GenAI 管道的文本 PII 编辑步骤。

C. Textract 提取文本；它不会编辑 PII。

D. Transcribe 将音频转换为文本；它不是文档的 PII 编辑器。