Skip to content
LLM Engineering Course
English
Esc
↑↓navigate↵open⌘Jpreview
On this page

Your route through the course

← Setup · Course home · Glossary

You do not need prior machine learning, calculus, or PyTorch: F04 teaches the derivative rules step by step, and you check each one by nudging a number. You do need a computer on which you can install Python, and time to try small programs. Unit 0 needs no Python knowledge; do it first, then the Python primer in F01 at your own pace. Unit 0 is a guided first success; it does not require you to understand the whole training system yet.

Choose the right starting point

Starting point Route
New to programming and ML Setup → Unit 0 → diagnostic → F01–F06 → T01–T06 → I01 → E01–E04 → C01–C04 → A01–A04 → Capstone
Python experience, new to ML Same route; demonstrate F01 skills in the diagnostic before skipping it.
Godot-RL course experience Unit 0 → diagnostic → B01 bridge → missing foundations → common F06 gate → T01–T06 → I01 → E01–E04 → C01–C04 → A01–A04 → Capstone.
Python and one LLM call, wants retrieval first Setup → R01 → R02, then the core route above.
Did Unit 0, wants to build an agent next Setup → Unit 0 → AG01 → AG02 → AG03, then the core route above.

Time, estimated from the budgets the pages state: setup, Unit 0 and the diagnostic take about 2–3 hours, T01–T06 about 8–10 hours, E01–E04 about 4–5 hours and AG01–AG03 about 3.5–5 hours. Most other lessons state no budget yet, so there is no honest total.

The diagnostic is not a barrier. An unanswered question means “learn this next”. You can always choose the full lesson instead of a shortcut.

What each stage adds

This table is a map; you don’t need to understand it yet.

Stage Question you will be able to answer Check before continuing
Unit 0 What changes when I increase one next-character score? Predict, run and explain your own intervention; keep the JSON report. Troubleshooting if a check fails.
Diagnostic Which foundations can I skip safely, and where do I re-enter if one fails? Common gate: one batched shape, one stable cross-entropy value, one backward/update cycle, gradient clearing explained, seeded reproduction.
B01 bridge Which Godot-RL skills transfer directly to language models? Complete the transfer challenges, then take the shared F06 checkpoint.
F01–F06 How does a program calculate an error and update numbers to reduce it? F06 pass rubric: own checkpoint.py, gradient check below 1e-6, wrong-sign diagnosis, two CPU runs within 1e-7.
T01 bigram Can the previous token predict the next? Recreate the data → loss → update → validation → sampling loop and explain a baseline comparison.
T02 tokenizer How does text become integers without losing Unicode? Roundtrip and edge-case tests pass; discuss the vocabulary/sequence-length/OOV trade-off.
T03 pipeline How do documents become batches of next-token targets? Redraw the chain text → batch → embeddings and explain each tensor, boundary and mask.
T04 attention Which earlier positions contribute to this prediction? Hand calculation of one attention row, heatmap and causality test in the report.
T05 Mini-GPT How do these pieces become a trainable decoder? Tiny-batch overfit, exact reload; trace every shape and separate the attention mask from loss masking.
T06 core gate Can I implement, diagnose and explain a decoder independently? At least 8/10 including all coding and defect points, tiny-batch NLL below 0.1 and exact reload.
I01 decoding How does one row of logits become one token, when does generation stop, and what does the cache keep? Temperature columns and masks by hand; name the stop rule and what the KV cache stores per block.
E01 training Can I interrupt a run and resume the same experiment? Match uninterrupted CPU parameters, derive the accumulation weights and explain clipping’s position.
E02 data quality Is my apparent improvement caused by contaminated data? Audit, repair and measure a deliberately leaky evaluation; write a data card.
E03 evaluation Does a change improve held-out predictions across seeds? Paired baseline/change report with uncertainty; reject the unchanged control; name a smoke-check blind spot.
E04 profiling What fits my time and memory budget? Derive the parameter count, compare an estimate to a measured CPU run and state the memory scope.
C01 strategy Is training the right next action for this company problem? Trace every chosen component to a requirement, a measured baseline gap and an acceptance test.
C02 company data How does a source snapshot become auditable training data? Explain snapshot, training record, development case and protected gold case.
C03 protected evaluation Why is a development score not gold evidence? Contamination report with no overlaps; explain iteration versus sealed proof.
C04 continued pretraining Does domain text help without losing the general baseline? State a promotion threshold, name the rollback artifact, explain why DAPT needs instruction evaluation.
A01 SFT & LoRA How do I teach a response while preserving the base? Loss alignment without a Trainer; zero initial LoRA delta, frozen base, safe reload and merge; report a failed gate.
A02 adaptation comparison Which adaptation does the evidence justify? Read a fair paired scorecard; separate storage savings from peak memory; reject an adapter that fails a gate.
A03 DPO How does a preference pair change a policy? One response log-probability sum, the clipped PPO ratio, proxy reward versus task success.
A04 RLHF bridge How does a reward model get exploited, and what would trigger a rollback? Describe one rollback monitor and the immutable reference needed to reproduce it.
Capstone Can another person reproduce my company-model decision? Decision record: CPU recipes from a clean checkout, baseline hashes, two rejected alternatives, the rollback step.
R01 RAG build How does a document become a cited answer without training? Write one evaluation question in your own file and report its Recall@5 for both presets.
R02 RAG quality Does my retrieval system refuse and cite correctly? Say, with numbers from your own run, which preset retrieves better, how often each refuses and what the judge cannot see.
AG01 agent loop Why does the loop stop, and what does the model see after a tool runs? Reference trace using both tools; the five Break reports; the five-run table; your own third tool.
AG02 native tool use What did the API take over, and what is still mine? Three reference reports on one question; the four Break reports; the five-run table with cost; a schema-based tool.
AG03 prompt injection What does each defence in the ladder actually buy, and why does removing a capability succeed where filtering text does not? Predict ranking next to your own three tables; one sentence on the biggest surprise; check_defenses.py’s output; the two Break outputs; the classifier’s catch and false-positive rates.

Completing these labs is evidence for the course rubrics; it is not an automatically issued certificate or a public-release approval.

How to study one lesson

  1. Predict: write one expected number, shape or direction before executing.
  2. Trace: follow a single example through the code. Name each axis.
  3. Build: implement the small exercise yourself in artifacts/my-work/.
  4. Break: change the suggested assumption and keep the error as evidence.
  5. Measure: compare to the unchanged baseline on the same data and seed.
  6. Explain: write why the result happened; “the test passed” is not a mechanism.

Start with the CPU command. Read a JSON report as a tree of named measurements; validation.loss means the loss field inside validation. Timing can change between runs even when the learned parameters match. Keep notes of commands, configuration, versions and results. Generated reports/weights stay in artifacts/.

When you get stuck

Difficulty Return to
def, brackets, loops or an exception are unfamiliar F01
You cannot say what [B,T,D] means F02
Scores, probabilities and loss seem interchangeable F03
You cannot predict an update’s direction F04
Backward pass or a gradient check is mysterious F05
An optimizer or seed feels like magic F06
You cannot say which positions an attention row reads T04
Temperature, top-k/top-p or the KV cache are unclear I01
A paired comparison, seed or held-out claim is unclear E03
A new abbreviation blocks reading Glossary

Try the exercise before opening its hint or reference. If the reference is still unclear, reduce it to one row and one number. Repeat only the missing checkpoint; you do not need to restart the course. The real beginner pilot is still required to validate pacing and comprehension with learners.

Was this page helpful?