Make a model follow the behavior you need.
Use one real tiny GPT to study training data, SFT, LoRA and QLoRA, DPO, distillation, quantization, and the decision to fine-tune at all.
Decide when fine-tuning is justified
Prepare clean instruction and preference data
Implement SFT, LoRA, and DPO objectives
Evaluate distillation and quantization tradeoffs
StoryByte, the roughly 1.09M-parameter GPT whose post-training runs and artifacts are reproducible through the companion lab.
LLM Foundations or equivalent knowledge. Build a Tiny LLM is the recommended hands-on preparation.
Open any module to review its exact sections. Progress and completion follow you through the course.
Topics include context, system controls, and weights, prompt / few-shot / RAG / fine-tune, InstructGPT: 1.3B beat 175B, and more.
Topics include training examples, special tokens, data failure modes, and more.
Topics include supervised fine-tuning, loss masking, overfitting, and more.
Topics include low-rank adapters, rank & alpha, merging, and more.
Topics include preference pairs, RLHF, DPO, and more.
Topics include deploy math, soft targets, distillation temperature, and more.
Topics include merge & export, int8 quantization, serving math, and more.
Topics include the method ladder, decision tree, known limits, and more.
AI engineers evaluating fine-tuning
ML practitioners moving into post-training
Teams comparing prompting, RAG, and weight updates
These are independent LDS courses that extend the closest skills. Choose the direction that matches what you want to build next.
Engineer what the model sees and returns.
Turn one LLM call into a dependable service.
Replace “it feels better” with evidence.
The first module establishes the language and example used throughout the rest of the course.
