Instructions: To add the content below to your course, use the following HTML code. Here is a link for those who need a refresher on using the HTML view in the Rich Content Editor. Additionally, to accurately replicate this page, you must download, save, and upload the images (Image 1) to your files, then link them to your page so your students can view the content. Here is a link for those who need a refresher on uploading and embedding an image in the Rich Content Editor.

Overview:

Machine learning models (ML) trained on large public web datasets learn by analyzing enormous amounts of openly available information, identifying patterns, and adjusting themselves through repeated prediction and correction — then being tested, refined, and monitored for quality and safety.

Machine learning training pipeline diagram

What Are Large Public Web Datasets?

Large public web datasets are massive collections of information sourced from publicly accessible content, such as:

Websites and articles
Online forums and discussions
Public documents and reports
Open-source code repositories
Public images and videos
Educational resources

Note: The data used is sourced from publicly available information or appropriately licensed materials for research and training purposes. While not all public data is considered ethically or legally suitable for training, this remains a topic of ongoing debate in AI ethics.

Step 1: Data Collection

Engineers gather large volumes of public data using automated systems. The goal is diversity — different topics, writing styles, languages, and formats. The purpose is to expose the model to a broad view of how information appears in the real world.

Step 2: Data Cleaning and Filtering

Raw web data is often disorganized; therefore, it needs to be processed to clean it. The goal of cleaning and filtering the data is to improve its quality and reduce noise, enabling the model to learn functional patterns.

Removing duplicates
Spam and low-quality content
Harmful or unsafe material
Corrupted or unreadable data

Step 3: Formatting and Structuring

Raw data is converted into machine-readable input formats that the model can process:

Text is tokenized (broken into smaller pieces)
Images are resized and normalized
Code and documents are structured

Step 4: Model Training

The model is exposed to data in large batches. The purpose is to gradually build a statistical understanding of patterns, language, and relationships. It learns by:

Making predictions (next word, label, output, etc.)
Comparing predictions to the actual data
Adjusting internal parameters to reduce errors
Training occurs on large Graphics Processing Units (GPUs) and Tensor Processing Units (TPUs), which operate in parallel to reduce training time from months to days.

Step 5: Evaluation and Testing

After training, the model is tested on separate datasets that it hasn’t seen before. The purpose is to ensure the model performs well beyond the training data. Engineers check for:

Accuracy testing
Bias and fairness checks
Stress testing for edge cases

Step 6: Fine-Tuning and Deployment

The base model may be refined with more focused datasets or human feedback. The purpose is to ensure the model is more helpful, aligned, and task-specific. Such as:

Moving the trained model to a production environment
Making it available to real users or systems
Letting it make predictions on new, live data
Monitoring performance and retraining as necessary
Connecting it through an Application Programming Interface (API), app, or service

- In simpler terms, an API is the bridge that lets apps and systems send data to a deployed ML model and receive predictions back.
  - Think of it like a server at a restaurant. You (the app) don’t go into the kitchen (the model). You tell the server (API) what you want. The server brings back the results. You don’t need to know how the kitchen works — just how to place the order.

The content on this page was created using Microsoft Copilot.

Microsoft. (2026, January 29). How Machine Learning Models Are Trained [Generative AI chat]. Copilot. https://m365.cloud.microsoft/chat/?fromCode=cmcv2&redirectId=FA22A70DFF5E41509CDDDC0A1AA19374&internalredirect=CCM&auth=2