Skip to main content

Large language models

From a raw pretrained model to a customer-support assistant in production. Ten modules, one running project, and every decision — open or proprietary, fine-tune or not, self-host or API — argued from measurements you can reproduce.

Course duration: 9h​

What you will learn​

  • Read scaling laws and know when a bigger model stops paying for itself
  • Turn a base model into an instruction-following assistant with supervised fine-tuning
  • Align outputs with human preferences using RLHF and its cheaper alternative DPO
  • Tune decoding — temperature, top-k, top-p — for each user-facing behaviour
  • Manage a finite context window: sliding summaries, external memory, when to reach for RAG
  • Diagnose the causes of hallucinations and pick the remedy that actually cuts them
  • Quantize a 7 to 8 billion parameter model to 4 bits and serve it with vLLM or llama.cpp
  • Build a custom evaluation set that catches contamination and matches your business
  • Compute the true cost per request and choose between API, self-hosting and a router

Prerequisites​

  • Transformer architecture (course 12)
  • Natural language processing fundamentals (course 13)

Course modules​

  1. What scaling really changes
  2. Pretraining: data, tokens and scaling laws
  3. Supervised instruction tuning
  4. Alignment: RLHF and preference methods
  5. Decoding: temperature, top-k, top-p
  6. Context window and memory management
  7. Hallucinations: causes and remedies
  8. Quantization and cost-efficient serving
  9. Evaluation: benchmarks and human judgment
  10. Costs, latency and architecture choices

The running project​

Every module contributes one decision to the same case study: deploying a customer-support assistant for a mid-sized company, in the language of the front. The reference open model manipulated throughout the code is a 7 to 8 billion parameter model under a permissive licence, served with vLLM or llama.cpp depending on the module. By module 10 you have a defensible answer to the questions that decide the project: what to fine-tune, what to retrieve, what to serve, what to pay.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number can be verified by anyone on the platform.

Free courses, by contrast, end with a 5-question quiz and a certificate preview, without certification.