Azure DP-100 key terms

The top-level Azure resource that groups everything a team uses for ML: compute, datastores, data assets, environments, jobs, models, and endpoints. Access control and linked services such as Storage and Key Vault are configured here.

A managed single-node virtual machine created by the workspace for interactive work such as notebooks and small experiments. It starts and stops on demand and comes with common ML frameworks preinstalled.

An autoscaling group of CPU or GPU nodes that the workspace starts on demand to run training jobs and pipelines. It scales from zero to a maximum node count and shuts nodes down when idle.

A saved connection between the workspace and a storage service such as Blob Storage, Data Lake Storage, or Azure SQL. It holds the connection details so jobs and data assets can read data without embedding credentials.

A named, versioned reference to data (a file, folder, or table) registered in the workspace. Jobs consume a specific version, which keeps experiments reproducible and preserves lineage.

A reproducible description of the software stack for a job or deployment: base image, Python packages, and variables. Environments are versioned so training and scoring use identical dependencies.

An Azure AD identity created and rotated by Azure and attached to a resource such as a workspace or compute cluster. Code uses it to authenticate to storage or Key Vault without storing secrets.

A network interface that connects a virtual network to an Azure service over a private IP address. For Azure ML it keeps workspace and storage traffic off the public internet.

A named container in the workspace that groups related runs so their metrics, logs, and artifacts can be compared side by side. Every job submission is recorded under an experiment.

A job type that runs a script or command on a compute target inside a specified environment. It is defined in YAML with a code location, inputs, outputs, and arguments.

A feature that tries many algorithms and preprocessing options for classification, regression, forecasting, or vision tasks and returns the best scoring models, with optional featurization and explanations.

A job that runs many trials over a search space of hyperparameters using a sampling method such as random or Bayesian, tracking each trial's metrics to find the best configuration.

A rule in a sweep job that stops weak trials before they finish, for example bandit or median stopping, so compute is spent on promising hyperparameter combinations.

An open-source logging API that records parameters, metrics, and artifacts for each run. Azure ML accepts these logs so runs can be compared in the studio or in an external tracking server.

A change over time in the distribution of model inputs compared with the training data. Monitors compare recent production data to a baseline and alert when drift may hurt predictions.

A reusable, versioned pipeline step with typed inputs and outputs plus its own code and environment. Components let teams assemble pipelines from building blocks shared across projects.

A workflow of components connected by data dependencies and submitted as a single job. Pipelines make multi-step training repeatable and can run each step on different compute.

A drag-and-drop canvas in Azure Machine Learning studio for building pipelines from prebuilt modules without writing code. Finished graphs can be submitted as jobs or published as endpoints.

A model artifact plus metadata stored and versioned in the workspace or registry. Registration gives the model a name and version so deployments and evaluations reference an immutable file.

A real-time REST endpoint hosted by Azure that serves a model through one or more deployments. It handles authentication, autoscaling, and splitting traffic between deployments.

A running instance of a model behind an endpoint, with its own environment, scoring script, and compute capacity. Multiple deployments on one endpoint enable safe rollout and traffic allocation.

An endpoint that scores large datasets asynchronously and writes results to a datastore. It is invoked with a job-style request that points at input data instead of a single record.

The Python file a deployment runs, containing an init function that loads the model and a run function that receives request data and returns predictions.

Continuous collection of signals from a deployed model, such as input data drift, prediction distribution, and latency, with alerts when behavior moves outside expected bounds.

A standard folder layout holding model files, an MLmodel metadata file, and an optional signature describing inputs and outputs. Azure ML can deploy this format without a custom scoring script.

Techniques that quantify how much each feature contributed to a prediction. SHAP values are the common approach, and Azure ML can generate explanations for trained and AutoML models.

A single view combining error analysis, fairness metrics, feature importance, and counterfactual examples for a model. Teams use it to audit behavior before and after deployment.

Writing and refining the instructions, examples, and context given to a language model so it produces the desired output, including few-shot examples and output-format rules.

A pattern that retrieves relevant documents from a search index and inserts them into the prompt, so the model answers from your data rather than only its training knowledge.

A dense numeric vector representing text or another item, positioned so that items with similar meaning sit close together. Embeddings power semantic search, clustering, and RAG retrieval.

A data structure that stores embeddings and supports fast approximate nearest-neighbor search, letting a query retrieve the most semantically similar documents or passages.

Additional training of a pretrained model on your own labeled examples so it adapts to a domain, style, or task. Unlike prompting and RAG, it changes the model's weights.

A parameter-efficient fine-tuning method that trains small low-rank matrices attached to existing layers while the base model stays frozen, cutting GPU memory and cost sharply.

The unit of text a language model reads and generates, roughly a word or word piece. Token counts drive context-window limits and are the basis for usage pricing.

A sampling setting that controls output randomness. Low values make responses focused and repeatable; high values make them more varied and creative.

An evaluation measure of whether a generated answer is supported by the source context supplied to the model, rather than being unsupported or hallucinated.

A development tool for building, testing, and deploying LLM applications as flows of connected nodes (prompts, Python code, and retrieval steps) with evaluation and tracing built in.

A screening layer that inspects model inputs and outputs for harmful categories such as hate, violence, sexual content, and self-harm, blocking or flagging them before users see them.