AI310: AI Technology Professional UPDATE

Agenda:

• Understanding Transformers: From Attention to Inference

  • From sequence models to Transformers: The role of attention and the problems it solves

  • Tokenization, embeddings, and positional information as model inputs

  • Self-attention step-by-step: Queries, keys, values, scores, softmax, and weighted representations

  • Masking and multi-head attention: Parallel perspectives on contextual relationships

  • Transformer blocks: Attention, feed-forward networks, residual connections, and layer normalization

  • Distinguishing model families: Encoder-only, encoder-decoder, and decoder-only

  • Classifying BERT, T5, and GPT: pretraining goals, strengths, and typical tasks

  • Understanding inference: Next-token prediction, decoding strategies, context windows, and KV cache

  • Hands-on Lab: Explore tokenization and embeddings, trace attention processes in the notebook, and compare pretrained model families based on typical tasks

    
    

• Adapting, evaluating, and effectively deploying models

  • Starting with typical LLM tasks: sentiment analysis, extraction, translation, summarization, and text generation

  • Practical use of pre-trained models via model libraries or APIs

  • The customization ladder: prompting and in-context learning, tool use and agent-based workflows, RAG, task-specific models, and fine-tuning

  • Fine-tuning variants: supervised fine-tuning, instruction tuning, full fine-tuning, and parameter-efficient fine-tuning with LoRA (with live demo)

  • Decision models instead of text generation: classification, routing, and evaluation with calibrated probabilities, using Jev from TypeSafe AI as an example

  • Planning training data: quality, labeling, data splitting, leakage, and overfitting

  • Evaluating models: task-specific metrics, generative quality, LLM-as-a-Judge, human review, and the limitations of automated evaluations

  • Making operational decisions: model size, reasoning vs. standard models, quantization, computational requirements, latency, and costs

  • Considering risks: hallucinations, bias, data privacy, prompt injection, and model licenses

    Hands-on Lab: Develop and justify a viable adaptation and evaluation strategy for your own use case

Objectives:

Upon completion of the workshop AI300 AI Technology Implementer participants will be able to:

• explain self-attention, queries, keys, values, and multi-head attention in a clear and technically accurate manner

• classify the key components of a Transformer model and understand their functions

• distinguish between encoder-only, encoder-decoder, and decoder-only models based on their strengths and tasks

• explain tokenization, pretraining, context windows, and autoregressive inference

• make informed choices between prompting, tool use, RAG, task-specific models, and fine-tuning

• assess when a calibrated decision model for classification and evaluation is a better choice than a generative LLM

• classify instruction tuning, full fine-tuning, and PEFT methods such as LoRA based on data, computational, and operational requirements

• plan a robust data and evaluation strategy for model adaptation

• Realistically evaluate the technical limitations, costs, and risks associated with an LLM use case

Target audience:

The training AI310 AI Technology Professional is targeted at:

• Data Scientists, ML Engineers, AI Engineers (beginner to intermediate)

• Data/AI Architects, Tech Leads, Engineering Leads in a data/AI context

• Product/project roles with a technical focus who want to support LLM decisions in a well-founded way

Prerequisites:

To participate in the course AI310 AI Technology Professional the following prior knowledge is required:

• Basic knowledge in Python

• A basic understanding of key machine learning terms: training, validation and test data, features and labels, overfitting

• Basic knowledge of vectors and matrices is helpful, but a brief refresher on the attention mechanism will be provided

Description:

The course AI310 AI Technology Professional provides a solid technical understanding of modern language models based on the Transformer architecture. Starting from tokenization, embeddings and self-attention, participants explore the interplay between queries, keys, and values, multi-head attention, and the other components of a Transformer block.

Building on this, encoder-only, encoder-decoder and decoder-only models are systematically compared. Participants learn to classify BERT, T5, GPT and modern LLMs based on their pretraining objectives, inference logic, strengths and limitations.

The second part of the course focuses on practical decision-making: When is prompting sufficient, when is RAG appropriate and when is fine-tuning worthwhile? In addition to full fine-tuning, parameter-efficient approaches such as LoRA are covered. For classification and evaluation tasks, participants also learn about decision models that provide calibrated probabilities instead of text. In notebook exercises and model comparisons, participants apply what they have learned and finally transfer it to their own use case, from data planning through evaluation to operations.