AWS AIF-C01 · AI 真题库（452 题） · 第 02 份

Evaluating models using a human workforce and custom prompt datasets ensures that the model generates responses in the style that aligns with the company's preferences. The other options either do not provide direct feedback on style preferences or are not specific enough for determining suitability basedonemployeepreferences.

使用人工团队和定制的提示数据集来评估模型，能够确保模型生成的回复风格符合公司的偏好。其他选项 要么无法直接反馈风格偏好，要么不够具体，无法根据员工偏好来确定其适用性。

Copying content from generative Al to write essays without proper attribution constitutes plagiarism, which is a key challenge of responsible generative Al. The other options are unrelated to this specific issue.

从生成式人工智能复制内容来撰写文章而不注明出处属于抄袭，这是负责任的生成式人工智能面临的一个 关键挑战。其他选项写此特定尚题无关。位法长

Amazon EC2 Trn series instances (powered by AWs Trainium chips) are designed to provide efficient and environmentally friendly training of large machine learning models. They are optimized for energy efficiency, which reduces the environmental impact of the training process. The other instance types are not specifically optimized for minimizing environmental effects during training.

亚马逊EC2Trn系列实例（由AWSTrainium芯片提供支持）旨在为大型机器学习模型的训练提供高效且环 保的解决方案。它们针对能源效率进行了优化，从而降低了训练过程对环境的影响。其他实例类型并未专门 针对训练过程中的环境影响最小化进行优化。

Guardrails for Amazon Bedrock can help ensure that the output generated by Amazon Bedrock is appropriate for children. Guardrails are used to apply content moderation, guidelines, and ensure safety by filtering potentially harmful or inappropriate content, which is essential when building an interactive application for children.

亚马逊Bedrock的防护栏可帮助确保由亚马逊Bedrock生成的内容适合儿童。防护栏用于实施内容审核、遵 循相关准则，并通过过滤潜在有害或不适当的内容来保障安全，这对于为儿童开发互动应用程序至关重要。

Amazon SageMaker Canvas is a no-code tool that allows users to build ML models and make predictions without requiring programming knowledge. It is ideal for users with no coding experience, providing an easy interface for importing data and generating predictive models. The other options require more technical expertise or are not designed for no-code model building.

亚马逊SageMakerCanvas是一款无需代码的工具，能让用户在无需编程知识的情况下构建机器学习模型并 进行预测。它非常适合没有编程经验的用户，提供了一个简便的界面用于导入数据和生成预测模型。其他 选项则需要更多的技术专长，或者并非专为无需代码的模型构建而设计。

Sampling bias occurs when the training data is not representative of the overall population, leading to disproportionate flagging of specific groups. In this case, the model may have been trained on biased data that did not adequately represent all ethnic groups, resulting in skewed predictions. The other types of bias do not directly apply to the selection of training data or its representativeness.

当训练数据不能代表总体人群时，就会出现抽样偏差，导致对特定群体的标记不均衡。在这种情况下，模型 可能是基于有偏差的数据进行训练的，这些数据未能充分代表所有种族群体，从而导致预测结果出现偏差。 其他类型的偏差并不直接适用于训练数据的选择或其代表性。

Reinforcement learning allows the chatbot to learn from interactions by receiving rewards for positive customer feedback, which helps the model self-improve over time. The other options do not directly provide a mechanism for continuous self-improvement based on interactions.

强化学习使聊天机器人能够通过接收积极客户反馈的奖励从交互中学习，这有助于模型随着时间的推 移不断自我改进。其他选项没有直接提供基于交互进行持续自我改进的机制。

A confusion matrix provides detailed insights into the performance of a classification model by showing the true positives, false positives, true negatives, and false negatives. This metric helps evaluate how well the model classifies the different types of materials in images. The other metrics are not as suitable for evaluating a classification model.

混淆矩阵通过展示真阳性、假阳性、真阴性和假阴性，为分类模型的性能提供了详细的见解。该指标有助于 评估模型对图像中不同材料类型的分类效果。其他指标则不太适合用于评估分类模型。

Implementing moderation APls can help filter and block inappropriate or unwanted images before they are returned by the chatbot. The other options do not directly address ensuring that the chatbot avoids returning inappropriate images.

实现审核API可以帮助在聊天机器人返回之前过滤和屏蔽不适当或不需要的图片。其他选项并不能直接解 决确保聊天机器人不返回不适当图片的问题。

Enabling invocation logging in Amazon Bedrock allows the Al practitioner to monitor and store the input and output data for model invocations. The other options are not directly used for logging model invocationsin Amazon Bedrock.

在业马逊Bedrock中启用调用日志记录功能，可让AI实践者监控并存储模型调用的输入和输出数据。其 他选项并非直接用于在亚马逊Bedrock中记录模型调用。

Batch transform is ideal for processing large datasets that do not require real-time predictions. It allows the company to perform inference on multiple GBs of data efficiently without needing immediate results. The other options are more suitable for scenarios requiring real-time or near real-time access.

批处理转换非常适合处理那些不需要实时预测的大数据集。它能让公司高效地对数GB的数据进行推理，而 无需即时结果。其他选项则更适合需要实时或近实时访问的场景。

Embeddings are numerical representations of real-world objects and concepts that help Al and NLP models understand and work with textual information more effectively by capturing relationships and similarities between words or phrases. The other options do not describe this concept.

嵌入是现实世界中的对象和概念的数值表示形式，有助于AI和自然语言处理模型更有效地理解和处 理文本信息，通过捕捉单词或短语之间的关系和相似性。其他选项并未描述这一概念。

Domain adaptation fine-tuning allows the FM to better understand the complex scientific terms by training it with domain-specific data, improving its performance on such specialized content. The other options are either insufficient or not directly related to handling complex terminology effectively.

领域适应微调通过使用特定领域的数据对FM进行训练，使其能够更好地理解复杂的科学术语，从而提高其 在处理此类专业内容时的表现。其他选项要么不够充分，要么与有效处理复杂术语没有直接关系。

Decreasing the temperature value makes the model's output more deterministic and consistent by reducing randomness in response generation. The other adjustments do not directly ensure consistent responses.

降低温度值可减少响应生成中的随机性，从而使模型的输出更具确定性和一致性。其他调整并不能直接 确保响应的一致性。

Creating a custom Amazon Bedrock service role for each team with restricted access to only the team's customer data ensures compliance with the security policy, providing the necessary data segregation and access control. The other options do not effectively enforce team-specific access or may pose risks of broaderaccess than allowedby thepolicy.

为每个团队创建具有仅限访问该团队客户数据权限的自定义亚马逊Bedrock服务角色，可确保符合安 全策略，实现必要的数据隔离和访问控制。其他选项无法有效实施团队特定的访问权限，或者可能会 导致超出策略允许范围的更广泛访问权限。

Guardrails for Amazon Bedrock can be used to filter content and ensure that personal patient information is not included in model responses. Setting up Amazon CloudWatch alarms allows the company to receive notifications when policy violations occur. The other options are not specifically designed for filtering model output and monitoring policy compliance.

亚马逊Bedrock的防护栏可用于筛选内容，确保模型响应中不包含个人患者信息。设置亚马逊 CloudWatch警报可让公司收到政策违规的通知。其他选项并非专门用于筛选模型输出和监控政策合规 性。

Amazon Textract can extract text from PDF documents, making it suitable for converting resumes into plain text for further processing. The other services do not provide functionality to extract text from PDFs.

亚马逊Textract能够从PDF文档中提取文本，因此适合将简历转换为纯文本以进行进一步处理。其 他服务不具备从PDF中提取文宋的能

Adding a role description to the prompt is the simplest and most effective way to adjust the model's response style based on the user's age range. It requires minimal implementation effort and effectively tailors the output. The other options involve more complex processes, such as fine-tuning or additional reasoning steps.

在提示中添加角色描述是根据用户年龄范围调整模型响应风格最简单且最有效的方法。这种方法实 施起来几乎不费力，还能有效地定制输出内容。其他选项则涉及更复杂的流程，比如微调或额外的 推理步骤。

Evaluating a foundation model's accuracy by measuring its performance against a predefined benchmark dataset is the standard approach for assessing accuracy in image classification tasks. The other options do not provide an appropriate measure of classification accuracy.

通过将基础模型在预定义的基准数据集上的表现进行衡量来评估其准确性，这是评估图像分类任务准 确性的标准方法。其他选项并不能提供恰当的分类准确度衡量标准。

Include fairness metrics for model evaluation: Fairness metrics help ensure that the LLM is unbiased and treats all cases equitably, which is essential for responsible Al use. Modify the training data to mitigate bias: Adjusting the training data helps reduce any inherent bias that might exist, contributing to a more fair and responsible LLM. The other options are related to general model optimization but do not directly address responsible Al practices regarding potential harms like bias and fairness.

在模型评估中纳入公平性指标：公平性指标有助于确保大型语言模型不存在偏见，并公平对待所有情 况，这对宇负责任地使角人工智能至关董要。口，国司 修改训练数据以减轻偏差：调整训练数据有助于减少任何固有的偏差。 可能会存在，有助于打造一个更公平、更负责任的语言模型。 其他选项涉及的是通用模型优化，但并未直接针对有关潜在危害（如偏见和公平性）的负责任的人工 智能实践。

The company is currently in the exploratory data analysis (EDA) stage, which involves summarizing data through statistics, visualizations, and correlation matrices to understand the dataset before moving on to modeling. The other options are subsequent steps in the ML pipeline.

该公司目前处于探索性数据分析（EDA）阶段，这一阶段通过统计、可视化和相关矩阵来总结数据，以便 在进行建模之前了解数据集。其他选项是机器学习流程中的后续步骤。

BERT-based models are well-suited for natural language understanding tasks, including filling in missing words, because they use contextual information to predict missing tokens in a text. The other types of models are not designed for this type of text completion task.

基于BERT的模型非常适合自然语言理解任务，包括填补缺失的单词，因为它们利用上下文信息来预测文 本中缺失的标记。其他类型的模型并非为此类文本补全任务而设计。

Amazon Q in Amazon QuickSight allows users to ask questions in natural language and automatically generate graphs and visualizations to display insights, such as total sales for top-selling products. The other options do not provide the same functionality for generating visual analytics.

亚马逊Q是亚马逊QuickSight中的一项功能，它允许用户以自然语言提问，并自动生成图表和可视化 内容来展示分析结果，例如畅销产品的总销售额。其他选项不具备生成此类可视化分析的功能。

Few-shot learning involves providing the model with a few examples to help it understand how to perform the task. For intent detection, the company needs pairs of user messages and the correct user intents, which will help the LLM improve its accuracy in detecting user intents. The other options do not provide the necessary pairing for improving intent detection.

少样本学习涉及为模型提供少量示例，以帮助其了解如何执行任务。对于意图检测，公司需要用户消息与正 确用户意图的配对，这将有助于大语言模型提高其检测用户意图的准确性。其他选项无法提供改进意图检测 所需的配对。

Decreasing the number of tokens in the prompt reduces the amount of data being processed, thereby lowering the cost of using the model. Since the model is performing well, reducing the prompt size is a cost-effective way to maintain performance while lowering expenses. The other options either increase costs or are unrelated to prompt size.

减少提示中的标记数量会降低处理的数据量，从而降低使用模型的成本。由于模型运行良好，减小提示 大小是在保持性能的同时降低费用的一种经济有效的方法。其他选项要么会增加成本，要么与提示大小 无关。

Hallucination occurs when a large language model generates content that appears plausible and factual but is incorrect or fabricated. This is a common issue with LLMs. The other options do not describe this particular behavior.

幻觉是指大型语言模型生成的内容看似合理且真实，但实际上却是错误的或虚构的。这是大型语言模 型常见的一个问题。其他选项并不能描述这种特定的行为。

To ensure that the custom model does not generate responses based on confidential data, the best approach is to retrain the model without including the confidential data. This prevents the model from learning patterns associated with that sensitive information, thereby avoiding its use in inference. The other options do not address the root cause of the issueremoving confidential data from the training process.

为确保自定义模型不会基于机密数据生成响应，最佳方法是重新训练模型，且不包含机密数据。这 样可防止模型学习与该敏感信息相关的模式，从而避免在推理过程中使用这些信息。其他选项并未 解决根本问题一一即在训练过程中移除机密数据。

The BLEU (Bilingual Evaluation Understudy) score is a common metric used to evaluate the accuracy of machine translation by comparing the generated translation with reference translations. It is specifically designed for translation tasks, whereas the other metrics are not suitable for evaluating translation quality.

BLEU（双语评估替补）分数是一种常用的评估机器翻译准确性的指标，它通过将生成的译文与参考译 文进行比较来实现。该指标专门针对翻译任务而设计，而其他指标则不适合用于评估翻译质量。

Amazon Bedrock agents help automate repetitive tasks and orchestrate complex workflows, which is ideal for handling thousands of customer support inquiries efficiently. This helps reduce response times and improves productivity. The other options do not directly address automation and orchestration of tasks for customer support.

亚马逊Bedrock代理可帮助自动执行重复性任务并协调复杂的工作流程，这对于高效处理数干个客户支持咨 询非常理想。这有助于缩短响应时间并提高工作效率。其他选项并未直接针对客户支持任务的自动化和协 调进行处理。

Ongoing pre-training helps enhance a foundation model's performance by continuously updating it with new data, thereby improving its ability to generalize and perform well on different tasks. The other options do not directly relate to the benefits of ongoing pre-training.

持续的预训练有助于通过不断用新数据更新基础模型来提升其性能，从而增强其泛化能力，并使其在不同 任务中表现出色。其他选项与持续预训练带来的益处没有直接关联。

Tokens are the smallest units (e.g., words, subwords, or characters) that generative Al models use to process text. They form the basis of both the input given to and the output generated by the model. The other options do not accurately describe tokens in this context.

标记是生成式人工智能模型处理文本时使用的最小单位（例如单词、子词或字符）。它们构成了模型输 入和输出的基础。其他选项在此语境下不能准确描述标记。

Inference costs for large language models are typically driven by the number of tokens processed during input and output, as each token incurs computational resources. The other factors (temperature value, training data, and training time) do not directly impact inference costs.

大型语言模型的推理成本通常由输入和输出过程中处理的标记数量决定，因为每个标记都会消耗计算资源。 其他因素（温度值、训练数据和训练时间）不会直接影响推理成本。

Configuring Amazon SageMaker to use a VPC with an S3 endpoint ensures secure, direct, and managed data flow between Amazon S3 and SageMaker Studio notebooks. This setup avoids public internet exposure and maintains data integrity during transfers. The other options do not provide a solution for managing the data flow in this context.

将Amazon SageMaker配置为使用带有S3端点的VPC，可确保Amazon S3与SageMakerStudio笔记本之 间实现安全、直接且受管理的数据传输。这种设置避免了通过公共互联网进行数据传输，并在传输过程中 保持数据的完整性。其他选项在此情况下无法提供管理数据流的解决方案。

Amazon S3 is the most suitable AWS service for uploading and storing datasets used for validation purposes. It is highly scalable and integrated with Amazon Bedrock, allowing easy access to data for model validation. The other options do not provide the same level of integration or suitability for managing datasetsinthiscontext.

业马逊S3是用于上传和存储用于验证目的的数据集的最合适的AWS服务。它具有高度的可扩展性，并与 亚马逊Bedrock集成，便于模型验证时访问数据。其他选项在集成程度或在此类场景下管理数据集的适用 性方面均无法与之相比。

"Extracting the prompt template" is a type of prompting attack that involves directly exposing the configured behavior or the underlying system prompt of a large language model (LLM). This attack can reveal sensitive details about how the model operates, including its internal instructions or restrictions, which are typically not intended to be disclosed to the user. Such an exposure can compromise the security and reliability of the LLM by making it vulnerable to further exploitation or misuse.

“提取提示模板”是一种提示攻击类型，它涉及直接暴露大型语言模型（LLM）的配置行为或底层系统 提示。这种攻击可能会泄露有关模型运行方式的敏感细节，包括其内部指令或限制，而这些通常不打算 向用户披露。这种暴露会损害LLM的安全性和可靠性，使其容易受到进一步的利用或滥用。

According to the AWS Shared Responsibility Model, AWS manages the security of the cloud, including the infrastructure and services like Amazon Bedrock. Customers are responsible for security in the cloud, which encompasses protecting their data, managing access controls, and configuring security settings for their applications. Reference:https://aws.amazon.com/compliance/shared-responsibility-model/?nc1=h_ls

根据AWS共同责任模型，AWS负责云的安全，包括基础设施以及诸如亚马逊Bedrock之类的服务。客户则 负责云中的安全，这包括保护其数据、管理访问控制以及为其应用程序配置安全设置。 鑫麦：（htps:/awsamazicomcamRliance/shared-responsibility-model/?ncl=h_Is亚马逊网络

Automatic model evaluation is the strategy that allows the company to evaluate the LLMs with the least operational overhead. This method leverages automated tools and processes to assess the toxicity or quality of the generated output without the need for manual intervention or crowd-sourced input. By using pre-built evaluation metrics or toxicity detection models, the company can quickly and efficiently evaluate multiple models without the complexity and time required for human evaluations.

自动模型评估是一种策略，能让公司以最低的操作成本来评估大型语言模型。这种方法利用自动化工具和 流程来评估生成输出的毒性或质量，无需人工于预或众包输入。通过使用预先构建的评估指标或毒性检测 模型，公司能够快速高效地评估多个模型，而无需人工评估所需的复杂性和时间。

"Jailbreaking" refers to attempts to bypass or disable the built-in safety features and restrictions of a system, in this case, a foundation model (FM). This technique involves trying to circumvent the safeguards that prevent the model from generating harmful or unsafe content. Jailbreaking is often performed to exploit vulnerabilities in a model's filtering or safety protocols, making it a direct attempt to undermine its protections.

“越狱”指的是试图绕过或禁用系统内置的安全功能和限制，在这里指的是基础模型（FM）。这种技术涉 及试图规避防止模型生成有害或不安全内容的保护措施。越狱通常是为了利用模型过滤或安全协议中的漏洞， 从而直接破坏其保护机制。

Network isolation in Amazon SageMaker allows you to run training and inference jobs in an environment that does not have access to the internet. This helps ensure that the data and the model do not inadvertently access external resources, meeting regulatory compliance requirements for isolated environments.

在Amazon SageMaker中，网络隔离功能可让您在无法访问互联网的环境中运行训练和推理作业。这有助 于确保数据和模型不会意外访问外部资源，从而满足隔离环境的监管合规要求。

Amazon SageMaker Model Cards are designed to document the essential details about machine learning models, including their intended uses, training datasets, training parameters, evaluation metrics, and inference environment. This provides a centralized mechanism to store and audit the metadata of the models, which is ideal for the ML team's need to share and audit models effectively. Key benefits of Amazon SageMaker Model Cards: - Standardized documentation of models' characteristics and intended use cases. • Transparency and traceability for auditing purposes. •Integration with other AWS services for lifecycle management.

亚马逊SageMaker模型卡片旨在记录有关机器学习模型的基本信息，包括其预期用途、训练数据集、训练 参数、评估指标和推理环境。这提供了一个集中机制来存储和审核模型的元数据，非常适合机器学习团队 有效共享和审核模型的需求。 亚马逊SageMaker模型卡的关键优势： ：模型特征及预期使用场景的标准化文档记录。 . ：审计目的所需的透明度和可追溯性。 与其他AWS服务集成以讲行牛命周期管理。

Natural language processing (NLP) tools can be used to generate code from high-level descriptions or suggestions, which can greatly enhance software development productivity. By leveraging NLP models, developers can automate repetitive coding tasks, generate code snippets, or even complete blocks of code based on natural language inputs. This can speed up development and reduce errors.

自然语言处理（NLP）工具可用于根据高级描述或建议生成代码，这能极大地提高软件开发的生产力。 通过利用NLP模型，开发人员可以自动执行重复的编码任务、生成代码片段，甚至根据自然语言输入完 成代码块。这可以加快开发速度并减少错误。

The Amazon SageMaker DeepAR forecasting algorithm is specifically designed for forecasting scalar (one-dimensional) time series data using recurrent neural networks (RNNs). It excels when trained on datasets containing hundreds of related time series, enabling it to learn patterns across multiple series and provide accurate forecasts. Reference:

亚马逊SageMakerDeepAR预测算法专门用于使用循环神经网络（RNN）对标量（一维）时间序列数据进行 预测。当在包含数百个相关时间序列的数据集上进行训练时，它表现出色，能够学习多个序列中的模式，并 提供准确的预测。 参考:

In machine learning, class imbalance occurs when certain classes are underrepresented in the training dataset, leading to biased model predictions. To develop an unbiased model, it's crucial to assess the class distribution and adjust the training process to address any imbalances. This can be achieved through techniques such as oversampling the minority class, undersampling the majority class, or applying class weights to ensure the model treats all classes equitably.

在机器学习中，当训练数据集中某些类别代表性不足时，就会出现类别不平衡的情况，这会导致模型预测出 现偏差。为了开发出无偏的模型，评估类别分布并调整训练过程以解决任何不平衡问题至关重要。这可以通 过对少数类别进行过采样、对多数类别进行欠采样或应用类别权重来实现，以确保模型公平对待所有类别。

Adversarial prompting is a technique used to defend against prompt injection attacks by crafting inputs that are specifically designed to identify and neutralize malicious prompts. This approach involves generating prompts that can detect and mitigate adversarial inputs, thereby enhancing the robustness of language models against such attacks.

对抗性提示是一种用于防范提示注入攻击的技术，通过精心设计输入来识别并中和恶意提示。这种方法包括 生成能够检测和减轻对抗性输入的提示，从而增强语言模型抵御此类攻击的能力。

The F1 score is a metric that combines precision and recall into a single value, providing a balance between the two. It is particularly useful in evaluating models where there is an uneven class distribution, as it considers both false positives and false negatives. The F1 score is calculated as the harmonic mean of precision and recall: recsoxe ecsn+ec This metric ranges from 0 to 1, with 1 indicating perfect precision and recall. In the context of evaluating a fine-tuned large language model (LLM) for a help desk application, the F1 score is appropriate because it assesses the mode's ability to provide accurate and relevant responses, balancing the trade-off between precision (correctness of responses) and recall (completeness of relevant responses). Reference:

F1分数是一种将准确率和召回率结合为单一值的度量标准，它在两者之间实现了平衡。在评估类别分布不 均衡的模型时，该分数特别有用，因为它同时考虑了误报和漏报的情况。F1分数是通过准确率和召回率的 调和平均值来计算的： Fl-2x(PrecisionxRecall) F-2x(Precision+Recall) 该指标的取值范围为0到1，1表示精度和召回率均完美。在评估用于帮助台应用程序的微调大型语言模型 （LLM）时，F1分数是合适的，因为它衡量了模型提供准确且相关响应的能力，平衡了精度（响应的正确 性）和召回率（相关响应的完整性）之间的权衡。 参考:

In Stable Diffusion, the classifier-free guidance (CFG) scale parameter controls how closely the generated image adheres to the provided text prompt. By increasing the CFG scale, the model places more emphasis on the prompt, leading to images that more accurately reflect the specified details. However, it's important to balance this setting, as excessively high values can result in less diverse and potentially lower-quality images.

在稳定扩散模型中，无分类器引导（CFG）比例参数控制生成图像与所提供文本提示的贴合程度。通过 提高CFG比例，模型会更注重提示内容，从而生成更准确反映指定细节的图像。但需要注意的是，要平 衡好这一设置，因为过高的值可能会导致图像多样性降低，甚至质量下降。

Retrieval Augmented Generation (RAG) integrates external data sources with LLMs to produce accurate and contextually relevant outputs without the need for extensive retraining. By connecting the chatbot to the company's policy documents, RAG enables the model to retrieve pertinent information in real-time, ensuring responses are both accurate and up-to-date. This approach is cost-effective as it leverages existing data without the computational expenses associated with retraining or fine-tuning large models. Reference:

检索增强生成（RAG）将外部数据源与大型语言模型（LLM）相结合，能够生成准确且与上下文相关的 输出，而无需进行大量的重新训练。通过将聊天机器人与公司的政策文件相连接，RAG使模型能够实时 检索相关信息，确保响应既准确又及时。这种方法成本效益高，因为它利用了现有数据，而无需承担重新 训练或微调大型模型所带来的计算成本。 参考:

Amazon Q Developer is a tool designed to help users with minimal programming experience to work with AWS Glue. It provides a graphical user interface and simplifies the creation of data transformation and extraction workflows, allowing users to perform tasks like querying and working with data without needing deep coding skills.

亚马逊Q开发者是一款专为编程经验有限的用户设计的工具，旨在帮助他们使用AWSGlue。它提供了一 个图形用户界面，并简化了数据转换和提取工作流的创建，使用户能够执行诸如查询和处理数据之类的任 务，而无需具备深厚的编程技能。

The company is actively seeking to ensure that the image classification model is trained on a diverse dataset that includes insect bite photos from various genders, ethnicities, and geographic locations. This reflects the fairness principle of responsible Al, which emphasizes creating models that make unbiased decisions across all demographic groups. By including a diverse range of data, the company is aiming to prevent biases that could lead to inaccurate diagnoses or treatments for certain groups of people. Fairness ensures that Al systems do not discriminate based on race, gender, geography, or other characteristics.

该公司正在积极努力确保图像分类模型是在一个多样化的数据集上进行训练的，该数据集包含来自不同性 别、种族和地理位置的昆虫叮咬照片。这体现了负责任的人工智能的公平性原则，该原则强调要创建能在 所有人口群体中做出无偏见决策的模型。通过纳入多样化的数据，该公司旨在防止可能导致对某些人群诊 断或治疗不准确的偏见。 公平性确保人工智能系统不会基于种族、性别、地域或其他特征进行歧视。