Microsoft AI-300 Exam Prep
Operationalizing Machine Learning and Generative AI Solutions (Page 2 )

• Fine-tuning requires a properly formatted training dataset.
• For classification or instruction-style fine-tuning, examples are supplied in JSONL format.
• Each line represents one training example, typically containing the input prompt and the desired completion or label.
• The file must then be uploaded or made available to the Microsoft Foundry fine-tuning workflow.

• A: Prompt Flow is useful for prompt experimentation, but it is not required to start fine-tuning.
• B: You do not need to deploy the base model to an online endpoint before fine-tuning.
• C: Tracing supports observability and evaluation, but it does not prepare the fine-tuning job or dataset.

• Input feature distributions differ from training data ? Analyze dataset drift metrics

• Model accuracy drops without code changes ? Analyze dataset drift metrics

• Endpoint latency increases under load ? Investigate scaling and infrastructure metrics

• Sustainability
• Performance efficiency
• Reliability

• Proxy server: Acts as an intermediary for requests, but does not necessarily provide an encrypted connection to the corporate network.
• NGFW: A firewall with advanced inspection and filtering capabilities; it may support VPNs but is not the remote-access solution itself.
• Security zone: A logical network grouping based on trust or security requirements, not a method for remote access.

• Microsoft Foundry is an enterprise platform for developers, data scientists, and IT teams.
• It supports the full AI application lifecycle: selecting models, connecting enterprise data, developing, evaluating, deploying, monitoring, and scaling solutions.
• It can include visual tools and low-code experiences, but it is not primarily designed to enable business users to build complete generative AI solutions independently.

• B: You still need to select and configure an appropriate model.
• C: Business users may participate through visual tools, but they are not the primary target audience.
• D: Foundry includes low-code tools, but it is not fundamentally a low-code-only platform.

• A: Concerns task scheduling and avoiding missed runs, not directly sizing the warehouse.
• B: The amount of data in a stream alone does not reliably determine compute needs; the SQL processing is what matters.
• D: A multi-cluster warehouse addresses concurrency, not the compute size needed for one task execution.

• GPS coordinates
• Date and time the image was taken
• Camera or device information
• Image dimensions and software used

• Log data: Records generated by systems or applications, not normally embedded in image files.
• Encrypted data: Information intentionally encoded to prevent unauthorized access.
• Sensitive data: A broad classification, not the specific type of information being extracted.

• Each Case must follow defined support steps.
• Each step must be completed within a specified time limit.

• Case Escalation Rules: Escalate Cases based mainly on age or criteria; they do not model a complete sequence of service steps.
• Process Builder with Scheduled Actions: Can automate time-based actions but is not the standard framework for Case service-level commitments.
• Omni-Channel Routing: Routes work to agents based on capacity and priority; it does not track Case lifecycle deadlines.

• Remote and global work: Access is based on the approved device, not a specific office or geographic location.
• Seamless login: Users may not need to repeatedly enter credentials on a recognized device.
• Limited equipment: It uses devices employees already have rather than requiring separate hardware.

• Geotagging: Restricts access based on location and may prevent legitimate remote work.
• Smart cards: Require issuing and managing additional physical equipment.
• Time-based logins: Limit access by time and do not provide a seamless authentication experience.

• Customers may boycott V’s products.
• Supermarket customers may stop buying from V.
• Investors may view V as having higher legal, ethical and reputational risk.
• V could face fines, lawsuits, contract losses and higher costs.
• Expected future profits may fall.

• The table is large.
• Queries frequently filter or join on particular columns.
• The table’s natural data order does not support good micro-partition pruning.
• Clustering improves performance enough to justify its maintenance cost.

• Custom domain: miami.weyland.com requires a paid App Service tier; Free and Shared are unsuitable.
• Two instances: Standard supports scaling to multiple instances.
• SSL: Standard includes SSL support.
• 12 GB storage: Basic provides only about 10 GB, so it does not meet the requirement. Standard provides substantially more storage.
• Cost minimization: Standard is the lowest tier among the choices that satisfies all requirements.

• Disable default accounts: Default or unused accounts may have known usernames or predictable settings, making them targets. They should be disabled, removed, or securely configured.
• Remove unnecessary services: Every running service creates potential vulnerabilities. Removing services the server does not need reduces exploitable entry points.

• Add the server to the asset inventory: Important for asset management, but not a hardening step.
• Document default passwords: Default passwords should be changed or eliminated—not documented for continued use.
• Send server logs to the SIEM: Valuable for monitoring and detection, but generally performed as part of logging/monitoring configuration.
• Join the server to the corporate domain: May be necessary operationally, but does not itself harden the server.

• Firewalls and routers
• Servers and applications
• Endpoint devices
• Authentication systems
• Cloud services

• Centralized log collection
• Event correlation
• Dashboards and reporting
• Alerting
• Compliance support

• MFA: Protects sign-in and authentication.
• Azure Key Vault: Protects deployment secrets and credentials.
• Answer: No (B).

• B — AWS DataSync: Transfers data between storage systems. It is not a marketplace or integration service for third-party datasets.
• C — Kinesis Data Streams and CodeCommit: CodeCommit stores source code, not third-party analytical datasets. Kinesis is for streaming data, not retrieving datasets from repositories.
• D — Kinesis Data Streams and ECR: Amazon ECR stores container images, not third-party datasets.

Why A is the best fit

Why the other options are weaker

• B — Create alerts in Dynamics 365 Finance: Alerts notify users about changes; they do not calculate tax.
• C — Implement the Common Data Model: CDM standardizes data structures but does not provide tax-calculation logic.
• D — Run a Power BI report: Power BI is primarily for analysis and reporting. A report might display tax rates, but it would not normally calculate and apply tax to sales orders operationally.
• E — Write scripts and code: This could work, but it is less aligned with the requirement for a quick, cost-effective implementation because custom development requires more time and maintenance.

bash 
OTEL_SERVICE_NAME=foundry-agent

bash 
OTEL_SERVICE_NAME=langchain-service

python 
import os 


os.environ["OTEL_SERVICE_NAME"] = "langchain-service" 


from azure.ai.openai... # import your tracing components 
from azure.monitor.opentelemetry import configure_azure_monitor 


configure_azure_monitor( 
    connection_string=os.environ["APPLICATIONINSIGHTS_CONNECTION_STRING"] 
)

• Agent application: OTEL_SERVICE_NAME=foundry-agent
• LangChain application: OTEL_SERVICE_NAME=langchain-service

DAX 
Previous Year Sales = 
CALCULATE( 
    [Total Sales], 
    PARALLELPERIOD('DATE'[Month], -1, YEAR) 
)

• CALCULATE evaluates [Total Sales] after modifying the filter context.
• PARALLELPERIOD shifts the current date period. -1 moves it backward, and YEAR specifies a one-year shift.
• 'DATE'[Month] is the date column used for the time calculation. In a proper date table, this should be a real date column or an appropriate calendar column—not merely text such as "January".

• Orders contains order-level information, such as the order date.
• Order Line Items contains the products, quantities, and line amounts associated with each order.
• Merging them on the order key, such as OrderID, places the order date alongside each line item. This allows you to calculate total order value and analyze it over time.
• Product attributes can remain in the Products table and be related to the resulting line-item table through ProductID.

• Merging line items with products addresses product attributes but does not first provide the order-level date and value needed for time analysis.
• DAX calculated columns or lists add unnecessary model complexity.
• Calculating order counts is not the first requirement and does not combine the required order and line-item data.

• Firewall
• Intrusion prevention and detection
• Antivirus or antimalware
• Web and email filtering
• Virtual private networking (VPN)
• Sometimes content filtering and centralized reporting

• ?? ?? ?? ??
• ?? NFS ?? SMB ?? ??
• ?? ???? ???
• ?? IP ??? ???? ?? ???
• ??? ?? ???? ?????

• B: ????? NFS ? Snowball Edge? ?? ?? ? AWS? ?? ? S3
• D: ????? NFS ? ????/Direct Connect ? File Gateway ? S3

• AWS Network Firewall can inspect and block traffic across a VPC when traffic is routed through the firewall. A Drop action rule can deny traffic from the suspicious source IP.
• GuardDuty findings are sent to Security Hub.
• A Lambda function can be invoked when Security Hub receives a new GuardDuty finding, extract the source IP, and update the Network Firewall policy with a blocking rule.

• GuardDuty threat lists: Used to identify or increase the severity of findings; they do not directly block VPC traffic.
• AWS WAF: Protects supported web applications, such as CloudFront or Application Load Balancers. It does not provide general VPC-wide traffic blocking.
• GuardDuty suppression rules: Suppress or hide findings. They do not block the suspicious traffic.

DAX 
Sales % of Total = 
DIVIDE( 
    [Total Sales], 
    CALCULATE( 
        [Total Sales], 
        ALLSELECTED(Products) 
    ) 
)

DAX 
Total Sales = 
SUM(Sales[SalesAmount])

Why ALLSELECTED is required

• The numerator, [Total Sales], is evaluated for the current product.
• ALLSELECTED(Products) removes the current product’s visual filter when calculating the denominator.
• It preserves filters applied at the report, page, or slicer level.

Why not ALL(Products)?

DAX 
CALCULATE([Total Sales], ALL(Products))

• The load balancer provides the listener’s frontend IP address.
• The active SQL Server node must also respond to that same listener IP.
• Enabling Floating IP allows the listener IP to be preserved and handled correctly by the active replica.
• Without it, SQL connections may be sent to a node that cannot correctly claim the listener address.

• A health probe for the SQL Server port
• A load-balancing rule using the listener port
• Floating IP enabled on that rule

• Controlar o fluxo de dados entre S3 e SageMaker;
• Aplicar regras de segurança por meio da VPC;
• Reduzir a exposição dos dados;
• Permitir acesso privado ao bucket do S3.

• Amazon Inspector: avalia vulnerabilidades, mas não gerencia o fluxo de dados.
• Amazon Macie: ajuda a descobrir e proteger dados sensíveis no S3, mas não cria a conexão com o SageMaker.
• S3 Glacier Deep Archive: é uma classe de armazenamento para arquivamento, não uma solução de conectividade.

• Help the team identify the root causes of delays.
• Empower the team to improve its workflow and practices.
• Facilitate removal of impediments.
• Avoid directing the team toward a solution without its involvement.
• Protect quality and sustainable delivery rather than simply accelerating everything.

• A: An impediment log may help, but merely recording problems does not empower the team to solve them.
• B: Directing the team to solutions conflicts with self-management.
• C: Adding impediments to the product backlog may incorrectly treat internal obstacles as product features and delays their resolution.

• New investment: $536,500
• Resulting NPV: $0

• B: Developer Edition has limited user, storage, and data-volume capacity, making realistic community testing difficult.
• C: A Developer Edition org cannot contain sandboxes, so it cannot provide separate environments for development, integration, and UAT.
• E: A standalone Developer Edition org is not part of the production org’s deployment path, so change sets cannot be used to deploy from it directly to that production org. The team would need Metadata API, Salesforce CLI, or another deployment method.

• Set credit quotas for an account or individual virtual warehouses.
• Define monitoring intervals.
• Trigger actions when usage reaches thresholds, such as sending notifications or suspending a warehouse.
• Help control unexpected warehouse spending.

• Account Usage: Provides historical usage and billing data; it does not enforce credit limits.
• Warehouse Limit Parameter: Not the Snowflake object used for account-wide credit controls.
• Credit Consumption Tracker: Not a standard Snowflake object.

• Security controls
• Processes and procedures
• Compliance evidence
• Performance against contractual requirements

• Service-level agreement (SLA): Defines expected service performance, such as uptime and response times.
• Memorandum of agreement (MOA): Documents a general agreement between parties, but does not necessarily provide audit rights.
• Supply chain analysis: Identifies risks across suppliers and dependencies, but is not a contractual mechanism for reviewing vendor controls.

• Each team’s Bedrock workflow uses a separate IAM service role.
• The role’s Amazon S3 permissions can be limited to only that team’s customer folder or objects.
• This enforces least privilege and prevents one team’s Bedrock application from accessing another team’s data.
• Access is enforced by AWS permissions, not by asking users to provide a customer name in each request.

• A single role with broad S3 access does not provide team-level isolation.
• Asking teams to specify a customer name is not an access-control mechanism.
• Redacting data does not prevent unauthorized access.
• A bucket policy alone is insufficient unless it explicitly enforces the required team-role restrictions.

• Minimum and maximum values
• Number of distinct values
• NULL information
• Other metadata used for pruning

• A password cracker would normally be identified by behavior such as loading password hashes or attempting dictionary/brute-force attacks.
• The question describes a routine audit, not password recovery.
• Simply encountering a protected file during an audit is characteristic of a misconfigured or insufficiently privileged FIM.

• Restrict a port to a defined number of MAC addresses
• Use statically assigned or “sticky” MAC addresses
• Shut down or restrict the port when violations occur

• Load balancer: Distributes application traffic; it does not protect switch MAC tables.
• IPS: May detect suspicious network activity, but is not the primary control for limiting switch MAC entries.
• NGFW: Filters traffic at the network boundary; it does not directly prevent MAC table flooding on access switches.

• Reporting relationships
• Organizational structures
• Leadership styles
• Team maturity
• Individual and group behavior
• Communication and motivation

• Acquire Project Team: Focuses on obtaining the people needed for the project.
• Develop Project Team: Focuses on improving team skills, relationships, and performance.
• Manage Project Team: Focuses on tracking performance, resolving issues, and managing changes within the team.

• A class/teacher fact should contain one row per Class ID. It can support the count of classes and contain attributes such as Teacher First Name and Period Number.
• An attendance fact should contain one row per attendance event, such as Class ID + student + date. It supports attendance counts by day and monthly averages.

• Teacher First Name ? Teacher fact
• Period Number ? Attendance fact

• Load balancer: Distributes application traffic; it does not protect switch MAC tables.
• IPS: May detect some attack traffic, but it is not the primary control for switch-port MAC learning.
• NGFW: Protects at network boundaries but does not directly limit MAC addresses learned on access ports.

• On-premises resources, such as an existing database
• Cloud resources, such as Azure services or virtual machines
• Connectivity that allows the application to use both environments

• Public cloud: resources are entirely provided through a public cloud provider.
• Private cloud: resources are dedicated to one organization, typically on-premises or privately hosted.

• Advanced Persistent Threat (APT): A sustained, targeted attack campaign or attacker group—not a vulnerability.
• Trojan: Malware disguised as legitimate software.
• Stealthware: A general term for malware designed to avoid detection; it does not specifically mean an unknown vulnerability.

• Crowd-sourced evaluation: Requires recruiting and managing many evaluators.
• Human-worker evaluation: Provides human judgment but requires coordination, labeling instructions, and quality control.
• RLHF: Is a training and alignment process, not simply a low-overhead model comparison method.
• Automatic model evaluation: Quickly evaluates outputs using predefined toxicity or quality metrics with minimal manual effort.

• 10 minutes have elapsed, or
• 20 MB of accounting data has accumulated

• A user connects with their Google account.
• Google associates the user’s SSH key with that identity.
• IAM roles determine whether they can log in.
• Access can be tied to a specific employee for auditing.
• Removing the IAM role removes their login access.

• compute.osLogin: regular OS login access
• compute.osAdminLogin: OS login with administrative privileges, typically including sudo

text 
enable-oslogin=TRUE

• Creating and training machine-learning models
• Testing and evaluating models
• Deploying models as usable endpoints

• Economy: Obtaining inputs at the lowest appropriate cost.
• Efficiency: Making the best use of resources, such as staff, money, and equipment.
• Effectiveness: Achieving the intended objectives or outcomes.

• Structure requirements documents
• Draft user stories and acceptance criteria
• Create templates for functional and nonfunctional requirements
• Turn natural-language descriptions into organized documentation or implementation examples

• A: “Automatically generating detailed requirements documents” overstates Copilot’s role; it assists but does not independently determine complete requirements.
• C: /help provides guidance about Copilot features, not reliable requirements-conflict detection.
• D: Stakeholder communication and meetings remain human responsibilities.

• Render thousands of animation frames
• Generate product images in multiple sizes
• Apply filters or transformations to a large image collection
• Produce visual effects for a video

• File policy in Microsoft Defender for Cloud Apps: Can detect activity involving files, but it is not designed specifically to investigate risk from departing or deleted users.
• Access review policy in Microsoft Entra: Reviews whether users should retain access to resources; it does not detect historical document downloads.
• Alert policy in Microsoft Defender for Office 365: Primarily handles alerts for email and collaboration threats, not insider-risk investigations involving departed users.

• Accept: Acknowledging the risk and choosing not to take additional action.
• Transfer: Shifting the risk or financial consequences to another party, such as an insurer.
• Mitigate: Reducing the likelihood or impact through controls, such as patching or access restrictions.
• Avoid: Eliminating the activity that creates the risk.

xml 
<apex:page> 
    <apex:stylesheet value="{!URLFOR($Resource.MyStyles, 'styles/main.css')}" /> 
</apex:page>

• <apex:includeScript> is for JavaScript.
• <apex:includeStyle> is not a valid Visualforce component.
• <apex:require> is not the Visualforce tag used to load CSS.