UPSKILLS.ACADEMY.AI
FT-LLM-01 // Free forever

LLM Foundations.
How the models work.

A plain-language course on NLP, tokenization, embeddings, transformers, and attention — then everything that happens after pretraining: fine-tuning, RAG, agents, and evaluation. Built for product builders who want real intuition, not hand-waving.

Self-paced · 8 modules Beginner → Intermediate No math background required
Start module 01 All 26 lecturesAll free trainings

Curriculum

Eight modules.

Module 01

What language models actually are

Prediction machines, not knowledge bases. From n-grams to neural LMs, and why next-token prediction produces something that looks like reasoning.

  • Next-token prediction
  • Probability distributions over text
  • Why models hallucinate
Module 02

NLP basics before transformers

The problems the field was trying to solve: classification, tagging, translation, summarization — and the pipelines used before deep learning.

  • Bag-of-words and TF-IDF
  • Word2Vec and GloVe
  • RNNs, LSTMs, and their limits
Module 03

Tokenization

Models never see words. They see token IDs. How byte-pair encoding splits text, and why tokenization explains cost, context limits, and weird failures.

  • BPE and subword vocabularies
  • Tokens vs. words vs. characters
  • Cost and context-window math
Module 04

Embeddings and vector space

Turning tokens into vectors where meaning becomes geometry. Similarity, clustering, and the foundation of semantic search.

  • Vector representations
  • Cosine similarity
  • Embedding models in production
Module 05

Attention and the transformer

The 2017 idea that changed everything. Queries, keys, values, multi-head attention, positional encoding, and the full decoder block — explained visually.

  • Self-attention step by step
  • Multi-head attention
  • Residuals, layer norm, feed-forward
Module 06

Training: pretraining to alignment

How a raw model becomes a helpful assistant. Pretraining scale, supervised fine-tuning, RLHF/DPO, and what "alignment" actually changes.

  • Pretraining objectives
  • Instruction tuning
  • RLHF and preference optimization
Module 07

Inference, prompting, and context

Temperature, top-p, sampling, system prompts, and context engineering. Why the same model gives different answers, and how to control it.

  • Decoding parameters
  • Prompt patterns that work
  • Context windows and KV cache
Module 08

Extending the model: RAG, tools, agents

When to fine-tune vs. retrieve vs. give the model tools. Chunking, retrieval quality, function calling, and evaluating a system you can trust.

  • RAG pipeline design
  • Function calling and agents
  • Evals, benchmarks, and guardrails

What you will be able to do

Explain how a transformer generates text, layer by layer
Reason about tokens, cost, and context limits before you build
Choose between prompting, RAG, and fine-tuning with confidence
Debug hallucinations, drift, and bad retrieval like an engineer

Reference

Plain-English glossary.

Token
The smallest chunk of text a model reads or writes — usually part of a word.
Embedding
A list of numbers representing meaning, where similar text sits close together.
Attention
The mechanism that lets each token weigh every other token for context.
Parameter
A learned weight inside the network; billions of them encode the model's behavior.
Context window
How many tokens the model can consider at once, prompt plus response.
RAG
Retrieval-augmented generation: fetch relevant documents, then let the model answer with them.

Next step

Now build with it.

Take the theory into practice — build your own digital twin, step by step, using the models you just learned about.

Digital Twin course