Edor.ai Learning Centre
LLM Classroom: large language model foundations
Starting from the Transformer architecture, tokenisation, parameters and self-attention, then on to the three training stages, RAG, quantisation, on-premises deployment, gating, and where this is heading. Explained with everyday analogies — no maths or coding background needed.
28 lessons
about 374 minutes
For: all teachers, panel heads, IT coordinators; no coding background needed
Syllabus
01
Module 1: What is actually happening inside a model
Transformers, tokenisation, parameters, embeddings and self-attention, said plainly
- 01Why teachers need to understand LLMs, not just use themStarting from an awkward silence in a panel meeting, this lesson separates knowing how to operate AI from knowing how to judge it, and maps a reading path through all twenty-eight LLM lessons. 11 min· Level: Foundation
- 02The Transformer in plain language: how models append one token at a timeUsing a school radio live-caption desk as the analogy, this lesson explains how a Transformer processes input layer by layer and why answers are assembled token by token rather than composed in advance. 13 min· Level: Foundation
- 03Tokenisation: why Chinese uses more slots than EnglishUsing standard courier box sizes as the analogy, this lesson explains what a token is, why Chinese consumes more tokens, and what that means for context length and length control. 13 min· Level: Foundation
- 04What parameters store: what 10B to a trillion actually meansUsing a veteran teacher's marking instinct as the analogy, this lesson explains that parameters hold patterns rather than entries, and why parameter count is a poor leading procurement criterion. 12 min· Level: Foundation
- 05Embeddings, positional encoding and self-attentionThree fresh analogies — a museum floor plan, shot numbers on a storyboard and counterpoint in a score — unpack how a Transformer turns text into workable understanding and how that links to a school knowledge base. 14 min· Level: Intermediate
02
Module 2: How a model gets made
Pre-training, fine-tuning and alignment — and which step hallucination and bias grow out of
- 06Pre-training, self-supervised learning and data cleaningTwo fresh analogies — a karaoke fill-in-the-blank game and a school magazine editorial desk — explain how a model teaches itself in the first training stage and why data quality decides everything after. 13 min· Level: Foundation
- 07Fine-tuning, SFT and where instruction-following comes fromUsing the analogy of a radio host learning from call-in demonstrations, this lesson explains how supervised fine-tuning turns a model that merely continues text into a real assistant. 13 min· Level: Intermediate
- 08Alignment with RLHF and DPOUsing the analogy of inter-school debate adjudication and blind ranking, this lesson explains how alignment training works and why a school still needs guardrails of its own. 14 min· Level: Intermediate
- 09Where hallucination and bias come from, and how teachers can intercept themThis lesson traces hallucination to the generation mechanism, shows where bias enters teaching materials, and sets out five school-level checks that keep AI in the role of a drafting assistant. 15 min· Level: Foundation
- 10Knowledge cutoffs and when models stop updatingWhat a knowledge cutoff is, how to verify one in official documentation, which teaching tasks it affects most, and how a panel can track model versions without rebuilding every prompt from scratch. 13 min· Level: Foundation
03
Module 3: The ecosystem and where models come from
Hugging Face, small language models, and running one on the device itself
- 11Hugging Face: a public model warehouse teachers can verifyA public site where AI models are stored openly. Even without coding, a teacher can check who built a model, what licence it carries, and what limits its authors admit. This lesson walks through it using the analogy of a museum specimen cabinet you can read without touching. 11 min· Level: Foundation
- 12Model cards and file formats explainedThe pile of filenames and documentation on a model page follows a fixed structure. This lesson treats the model card as an equipment manual and GGUF as airline baggage tiers, so a teacher or IT coordinator can judge in five minutes whether a model belongs in school. 12 min· Level: Foundation
- 13Small language models for school workA school does not necessarily need the largest model. This lesson uses the analogy of a school minibus versus a chartered coach to explain why a few billion parameters covers most daily work, and when a flagship is still worth calling. 12 min· Level: Intermediate
04
Module 4: The concepts you genuinely need to build something
Prompts, knowledge bases and RAG, chunking, context caching, agents, MCP and frameworks
- 14Prompt engineering in one lesson: treating AI like a print shop that needs a job ticketThe same AI can produce a usable worksheet in five minutes or an hour of editing. The difference is whether you filled in five fields on the job ticket. This lesson explains prompting as a print-ordering workflow any teacher can run. 13 min· Level: Foundation
- 15The school knowledge base and retrieval-augmented generationA general AI has read the internet but not your school's homework policy for this term. Retrieval-augmented generation has it pull relevant passages from the school filing cabinet before answering, like a school office clerk checking internal circulars before replying to a parent. This lesson explains how, and why it beats trying to train the model on your school. 14 min· Level: Foundation
- 16Chunking, reranking and lost in the middlePaste an entire long document into an AI and it often misses the crucial clauses in the middle. This lesson uses the analogy of marking only assigned pages to explain a well-documented effect, and gives teachers four immediate remedies: chunking, overlap, reranking and putting critical instructions last. 15 min· Level: Intermediate
- 17Context caching and cost control: the same slides do not need reloading every timeAsk ten questions about one curriculum document and everything from the second question onwards can be faster and cheaper. This lesson explains context caching through leaving the same classroom slides on screen and changing only the discussion question, and shows how that settles prompt order. 12 min· Level: Intermediate
- 18AI agents in practice: tool calling, permissions, and why MCP matters for procurementA model cannot open your attendance register, but it can fill in a work order asking the system to look it up. This lesson covers tool calling, what separates an agent from a chatbot, how to set permissions, and why MCP affects the cost of changing provider later. 14 min· Level: Intermediate
- 19LangChain and LlamaIndex: what to ask when those names appear in a proposalVendor documents often say "built on LangChain and LlamaIndex". Teachers do not need to code. This lesson uses a front-of-house and backstage split to explain what frameworks connect, and offers four procurement questions. 12 min· Level: Hands-on
05
Module 5: On-premises deployment and frugal fine-tuning
LoRA / QLoRA, quantisation, and running a model on the school network with Ollama and LM Studio
- 20Fine-tuning school tone on one GPU: how LoRA, QLoRA and PEFT workYou need not retrain an entire large model. This lesson uses a "overlay arrangement on the original score" analogy for LoRA and QLoRA, and lists when retrieval beats fine-tuning. 15 min· Level: Intermediate
- 21From FP16 to INT4: how quantisation shrinks models for school hardwareQuantisation rounds stored model weights to coarser scales so the same seven-billion model fits a consumer graphics card. This lesson uses a school height-recording analogy, compares FP16, INT8 and INT4, and shows how to route tasks between on-campus and cloud builds. 14 min· Level: Intermediate
- 22Running models on campus with Ollama and LM StudioStarting from one existing machine in the IT room, this lesson walks through running a model on the campus network with Ollama and LM Studio, honest hardware sizing, concurrency limits, and which teaching tasks belong on campus. 15 min· Level: Hands-on
- 23On-device AI, Apple Intelligence and where data goesStarting from the AI features on a student's iPad, this lesson explains why on-device processing has a structural privacy advantage, where its capability ceiling sits, and how a school applies the same principle at campus scale. 13 min· Level: Foundation
06
Module 6: Safety, evaluation, and the limits you cannot design away
Gating, how much to trust LLM-as-a-Judge and leaderboards, hallucination and cost, ethics and copyright
- 24AI guardrails with Llama Guard, NeMo and enforced structureUsing a broadcast delay line as the analogy, this lesson unpacks how incoming messages are filtered, how outgoing content is reviewed, how formats are locked, and what each layer of Edor.ai's four-layer gating guards. 15 min· Level: Intermediate
- 25Jailbreaks, prompt injection and how a school defendsUsing a tampered laboratory SOP as the analogy, this lesson separates prompt injection from jailbreaking and sets out five defences a school can actually implement, in order of effectiveness. 14 min· Level: Intermediate
- 26Privacy, bias and copyright — practical AI ethics for schoolsSeparating privacy, bias and copyright into steps teachers can follow today, covering data minimisation, classroom bias spot-checks and graded academic integrity. 14 min· Level: Foundation
- 27Reading AI leaderboards — MMLU, GPQA and Chatbot ArenaWhat public benchmarks actually measure, where LLM-as-a-Judge skews results, and why thirty to fifty school-based questions beat any ranking for procurement. 14 min· Level: Intermediate
07
Module 7: Where this is heading
Multimodality, reasoning models and embodied AI