To preserve the current values of variables set in the notebook and maintain the state for the current instance, you should stop the current kernel.
Memory State Retention: In Azure Machine Learning Studio notebooks, stopping the notebook kernel halts execution while allowing the compute instance to continue running.
Avoiding State Loss: While stopping or resetting the entire compute instance will completely clear your active in-memory variables and notebook state, cleanly stopping the kernel natively preserves your current progress and data states within that active notebook session.
Reference:
https://learn.microsoft.com/en-us/azure/machine-learning/quickstart-create-resources

Explanation:
Box 1: Model Registry
Deploy resources from a pipeline.
To standardize MLOps practices with declarative infrastructure and automated pipelines, the best choice is to use a Model Registry via Infrastructure as Code (IaC) for your core setup, supplemented by Azure CLI inside your pipelines for operational tasks. Prompt Flow is an application development tool and should not be used to deploy infrastructure.
A Model Registry (like Azure Machine Learning registry or MLflow) is the standard architectural pattern for managing ML artifacts declaratively.
Declarative Tracking: It stores model versions, lineages, and environments as code-like configurations.
Pipeline Integration: CI/CD pipelines can natively fetch from or push to a registry using configuration files.
Environment Isolation: It allows you to promote the exact same model artifact across Dev, Staging, and Production environments without rebuilding.
Box 2: Bicep templates
Define Azure resources declaratively.
Bicep templates are an excellent choice for defining Azure resources declaratively in an automated MLOps pipeline.
Native Azure Integration: Bicep supports all Azure Machine Learning (Azure ML) resources immediately upon release.
Declarative Syntax: You define the desired end-state of your infrastructure without writing complex deployment scripts.
No State Management: Azure manages the state automatically, unlike Terraform which requires a remote state file.
Tooling Support: Deep integration with Azure Pipelines and GitHub Actions allows easy deployment validation.
Reference:
https://azure.microsoft.com/en-us/blog/new-azure-capabilities-to-simplify-deployment-and-management/

When labeled data is unavailable, traditional accuracy metrics cannot be computed. Data drift detection monitors changes in input data distribution, serving as a proxy for potential performance degradation. This allows early detection of issues before labeled data becomes available.

In Microsoft Foundry, when configuring and running a fine-tuning job for analyzing customer feedback (e.g., classifying support tickets), you should first enable tracing for all inference calls in the evaluation pipeline.
Tracing is a critical step in the "Evaluate" phase of the fine-tuning workflow, allowing you to capture input/output examples, identify the root cause of classification errors, monitor latency, and analyze model behavior before and after training.
Note:
To fine-tune a model and evaluate it effectively, you should follow this sequence:
1. Enable Tracing for Initial Baseline
Before you fine-tune, you should indeed enable tracing for all inference calls in your existing evaluation pipeline.
Purpose: This creates a clear "paper trail" of how the base model is currently failing.
Benefit: Tracing captures the exact inputs, outputs, and intermediate steps. By analyzing these traces, you can identify if the low accuracy is due to the model not following instructions, missing domain-specific jargon, or struggling with specific ticket categories.
Setup: You can enable Automatic Tracing (currently in preview) in Foundry to log these details to Application Insights without changing your code.
2. Configure and Run Fine-Tuning
Once you have analyzed the traces and prepared your labeled dataset, you can proceed with the fine-tuning job.
3. Continuous Evaluation
After the job completes, you must compare the fine-tuned model against your original traces Reference:
https://devblogs.microsoft.com/foundry/a-developers-guide-to-fine-tuning-gpt-4o-for-image-classification-on-azure-ai-foundry