AI & ML pathway: build a small language model, then generative AI.

adds: words and the next word

Build Your Own Small Language Model

What a language model does: predict the next word from probabilities. You compare simple n-gram models with transformer models, prepare a small dataset, then train and evaluate your first small language model, and think about using AI responsibly on a problem in your own community.

• next-word prediction
• n-gram models and their limits
• the machine learning pipeline
• training and evaluating a model

adds: a small model, training

Train a Small Language Model

A hands-on skill badge course: frame a real language-model problem, build a simple tokeniser, prepare a dataset for a transformer and run the training loop yourself.

• framing the problem
• a simple tokeniser
• preparing the dataset
• the training loop

adds: tokens

Represent Your Language Data

How text becomes numbers. You clean raw text, split it into characters, words and subword tokens, build a byte-pair encoding tokeniser and represent meaning with embeddings. You also ask whose voices the data leaves out, and document a dataset with a Data Card.

• preparing raw text
• tokenisation and BPE
• embeddings, vectors and matrices
• documenting data ethically

adds: a neural network

Design and Train Neural Networks

How models learn. You separate signal from noise, spot overfitting and underfitting, and use training, validation and test splits. You build multilayer perceptrons, tune them, and see gradients, backpropagation and stochastic gradient descent at work, training a model with Keras.

• generalisation and overfitting
• multilayer perceptrons
• gradients and backpropagation
• training with Keras

adds: attention

Discover the Transformer Architecture

What sits inside a modern language model: attention and how its weights are worked out, masked and multi-head attention, positional embeddings, transformer blocks and how the model picks its next token. You visualise attention and write it in code.

• attention weights
• masked and multi-head attention
• positional embeddings
• transformer blocks and decoding

adds: a foundation model

What generative AI is

Where generative AI fits: AI is the broad field, machine learning learns patterns from data, deep learning uses large neural networks, and generative AI uses them to produce new text, images or code. You see how a model reads text as tokens within a limited context window, and where businesses already use it, from drafting and summarising to customer support.

• AI, machine learning and deep learning
• how generative AI differs
• tokens and context windows
• business use cases

You practise spotting where generative AI fits a business problem, and where it does not.

• Introducing AI

  The basic ideas behind AI and machine learning.

• Introducing Generative AI

  How generative AI produces new text and images, and where businesses use it.

adds: the prompt

Asking well

Prompting is a skill. You give a model a clear role, task, context and format; show it a few examples so it follows a pattern (few-shot prompting); ask it to work through a problem step by step (chain-of-thought); and refine the prompt when the answer misses. You also see how prompts can be misused, and how to guard against it.

• prompt structure
• zero-shot and few-shot
• chain-of-thought
• refining and testing prompts
• prompt misuse

You practise writing and improving prompts for everyday business tasks.

• Using Prompts and Prompt Engineering

  Writing and refining instructions to get useful answers from a model.

adds: a choice of models

Choosing and adjusting a model

Foundation models differ in size, cost, speed and what they do well. You compare the options available on AWS, for example through Amazon Bedrock, and shape answers with inference settings such as temperature and maximum length. To answer from your own documents, retrieval-augmented generation (RAG) finds the relevant passages first; fine-tuning trains the model further, at more cost and effort. You learn when each is worth it.

• comparing foundation models
• Amazon Bedrock
• temperature and other inference settings
• retrieval-augmented generation
• fine-tuning trade-offs

You practise choosing a model and its settings for a task, and deciding between RAG and fine-tuning.

• Working with Foundation Models

  Choosing a foundation model and adjusting it for a task.

adds: guardrails

Using it responsibly

Generative AI can be confidently wrong (hallucination), can repeat bias from its training data and can expose data it should not. You learn to test for these risks, add guardrails that filter harmful or off-topic output, keep people in the loop for important decisions, protect data and control access, and follow the governance and compliance rules that apply to your organisation.

• hallucination
• bias and fairness
• guardrails
• data privacy and access control
• governance and compliance
• human review

You practise reviewing an AI use case for its risks and the controls it needs.

• Practicing Generative AI Responsibly

  Fairness, accuracy and the risks to watch for when AI writes for you.

• Considering Security, Governance, and Compliance for Generative AI

  Keeping data safe and following the rules when a business uses generative AI.

adds: the application around it

Building the application

How a generative AI application comes together: define the business problem and how success will be measured, choose the model, connect it to your data, design the prompts and test the answers, then deploy it and watch quality and cost once people use it.

• defining the problem
• model and data choices
• testing and evaluation
• deployment
• monitoring quality and cost

You practise planning a generative AI application from idea to launch.

• Implementing Generative AI Applications

  How a generative AI application is planned and built, piece by piece.

closes with

Next steps

How the course connects to AWS Certification and where to go next. The exam itself is not included.

• Bridging to Certification

  How the course connects to AWS Certification.