Azure DP-100 study plan
Azure DP-100 · Give each week a clear focus.
8 weeks · About 6 hours per week
Design and prepare a machine learning solution
Provision an Azure Machine Learning workspace with the CLI and connect to it from a notebook using the SDK
Register a datastore backed by blob storage and create a versioned tabular data asset from it
Build a curated environment from a conda YAML file and validate it on a compute cluster
Justify the choice of managed versus attached compute for a given training workload
Your checkpoint
A setup repo containing workspace/compute/environment scripts plus a notebook that reads a registered data asset on a compute cluster, with a one-page compute-and-storage decision note.
Design and prepare a machine learning solution & Explore data, and run experiments
Split a dataset into train, validation, and test partitions and register each as a separate data asset
Implement feature engineering and encoding transforms in a notebook and persist the transformed dataset as a new asset version
Profile the data for missing values, class imbalance, and leakage risks and record the findings
Explain how to reuse the data asset across environments without hard-coded paths
Your checkpoint
A feature-engineering notebook that outputs a registered, versioned training dataset plus a written data-validation checklist covering at least five checks.
Explore data, and run experiments
Log parameters, metrics, and artifacts for every run using MLflow inside Azure Machine Learning
Run at least three comparable experiments that vary algorithm or feature set under one chosen evaluation metric
Compare runs side by side and select a baseline with a stated justification
Identify which experiment artifacts you would promote to a registered model
Your checkpoint
A tracked experiment containing at least three logged runs and a results table that names the baseline model and the metric used to pick it.
Train and deploy models
Refactor notebook code into a parameterized training script that accepts arguments
Submit the script as a command job on a compute cluster and retrieve its outputs and logs
Configure a hyperparameter sweep with a sampling method and an early-termination policy
Register the best model with tags, metrics, and a description
Your checkpoint
A command job and a completed sweep recorded in the workspace, a registered best model, and a Markdown summary of the sweep trials and their metrics.
Train and deploy models
Write a scoring script and deploy the model to a managed online endpoint
Send a test request to the endpoint, inspect invocation logs, and split traffic across two deployments
Configure a batch endpoint and score a multi-row dataset through it
Define a retraining and monitoring trigger based on data drift or performance decay
Your checkpoint
A working online endpoint and a working batch endpoint, plus a short runbook describing traffic splitting, rollback, and the retraining trigger.
Optimize language models for AI applications
Author versioned prompt templates with system instructions and few-shot examples
Build a retrieval-augmented generation flow over a document index and inspect retrieved chunks for relevance
Compare retrieval-augmented prompting against fine-tuning on the same task and state which fits the constraints
Tune chunking strategy and retrieval depth against a small question set
Your checkpoint
A retrieval-augmented prototype notebook that answers questions over at least 20 documents, with prompt templates and chunking settings committed to the repo.
Optimize language models for AI applications & Train and deploy models
Define an evaluation set and score model outputs for groundedness, relevance, and coherence
Run the evaluation as a pipeline or job and log the scores for baseline and tuned variants
Deploy the stronger variant to an online endpoint with content filtering and token limits configured
Add timeout, retry, and cost controls to the client code that calls the endpoint
Your checkpoint
An evaluation report comparing baseline versus tuned prompt scores on the same eval set, and a deployed endpoint serving the better variant.
Design and prepare a machine learning solution & Explore data, and run experiments & Train and deploy models & Optimize language models for AI applications
Re-run the two labs you scored lowest on until you can complete them without notes
Complete a timed practice assessment under exam conditions and review every missed item
Consolidate SDK commands for workspace, data, job, sweep, and endpoint operations into a single reference sheet
List and close remaining gaps in under a week of study
Your checkpoint
A scored practice assessment with a remediation list of no more than five gaps and a completed one-page SDK command reference sheet.
Milestones
Week 2: From an empty workspace you can register a datastore, a versioned data asset, and a curated environment in under 20 minutes without consulting notes.
Week 4: You can go from a registered dataset to a registered model through a command job and a hyperparameter sweep, without using a notebook for training.
Week 6: You can explain and demonstrate when retrieval-augmented prompting beats fine-tuning, and you have a working prototype that cites retrieved chunks.
Week 8: Practice assessment score is at or above your target pass threshold, and every logged gap has an assigned remediation lab.