Advanced engineering
Engineering
Available on request
LLM Foundations: How Language Models Actually Work
Under the hood. Tokens to transformers to fine-tuning, with the mathematics kept honest but human.
For the engineer who has to defend an architecture decision, not repeat a vendor slide.
- Duration
- 7 days, 40 contact hours
- Cohort size
- 8 to 12 participants
- Delivery
- In-house, Online, Blended
- Languages
- Arabic, English
Who it is for
Engineers, data scientists, technical architects, researchers, and anyone who has to make real architecture decisions about LLMs.
What people leave able to do
- Explain what happens between a prompt and a response, precisely
- Choose models on architecture and evaluation evidence rather than marketing
- Understand context windows, attention and their real cost implications
- Decide between prompting, RAG, fine-tuning and training, and justify it
- Read and interpret a model card and a benchmark result critically
Modules
- 01 Tokenisation and why it breaks Arabic
- 02 Embeddings and vector space
- 03 The transformer architecture
- 04 Attention, properly explained
- 05 Pre-training and what models learn
- 06 Instruction tuning and RLHF
- 07 Inference, sampling and temperature
- 08 Context windows and their economics
- 09 Open versus closed weights
- 10 Fine-tuning approaches and LoRA
- 11 Benchmarks and how to read them
- 12 Multilingual and Arabic-language model behaviour
Prerequisites
Python, and comfort with basic linear algebra and probability.
Start with a short call
Fifteen minutes to understand what your team does and what you want to change. If training is not the right answer, I will say so.