Level 13 of 14 · about 4 weeks

Advanced topics

Go deeper where the evidence says it pays: tuning small models, running them efficiently and reading research.

Modules

1. 1

   When to fine-tune

   Prompt, retrieve, or tune: a decision with evidence and a break-even

   Common 3 parts · about 3 hours

2. 2

   LoRA and QLoRA

   Train a small low-rank update instead of the whole matrix, and store the base in 4 bits

   Common 3 parts · about 4 hours

3. 3

   Preference tuning

   Reward models, RLHF, DPO and the KL leash, tested on a toy policy

   Common 3 parts · about 4 hours

4. 4

   Reinforcement learning for reasoning and agents

   Verifiable rewards, group-relative advantages, sparse rewards and reward hacking

   Specialist 3 parts · about 5 hours

5. 5

   Dataset engineering

   Deduplication, cleaning, label checks, leakage, and the limits of synthetic data

   Specialist 3 parts · about 5 hours

6. 6

   Inference optimisation

   KV cache, memory-bound decoding, batching, speculative decoding and quantisation

   Specialist 3 parts · about 5 hours

7. 7

   Small and open models

   Sizing, self-host versus API economics, routing and cascades, and testing a model switch

   Specialist 3 parts · about 5 hours

8. 8

   Multimodal agents

   Seeing and acting on screens: DOM versus screenshots, grounding, cost and reliability

   Specialist 3 parts · about 5 hours

9. 9

   Reading research papers

   Extract the claim, the setup and the catch, and check how much a reported gain can be noise

   Specialist 3 parts · about 4 hours