03 section
Training and Adaptation
How models are built and then bent to a purpose — pretraining, SFT, LoRA, preference alignment, distillation, synthetic data, quantization — in the order those steps actually happen
What is in here
Most teams never pretrain, so the value here is knowing which adaptation lever fits which problem, and being able to say out loud why RAG beats fine-tuning for facts. The pages are short and dense; LoRA and quantization are the two most likely to come up in an interview. Straight through in numeric order.
01
3 min
Pretraining Basics
Causal language modeling, data curriculum, and the compute arithmetic behind a pretraining run — enough to reason about why frontier training costs what it does
trainingfundamentalsdata
02
3 min
Fine-Tuning Strategies
Fine-tuning teaches form, RAG supplies fact; the line between them, then SFT, continued pretraining, and the hyperparameters that actually move results
fine-tuningtrainingproduction
03
3 min
LoRA, QLoRA and PEFT
Low-rank adapters explained through the rank and alpha math, why QLoRA fits a 70B on one GPU, and how multi-LoRA serving swaps skills at runtime
fine-tuningquantizationserving
04
3 min
RLHF and DPO
Alignment from reward models and PPO through to DPO's simpler contrastive objective, plus the online methods and verifier-based RL that train reasoning models
alignmenttrainingreasoning
05
3 min
Knowledge Distillation
Transferring a large teacher's behavior into a small student — output versus feature distillation, self-distillation from proof, and distilling with quantization in mind
distillationtrainingmodel-selection
06
3 min
Synthetic Data Generation
Human text ran out, so training data is now manufactured: Evol-Instruct, RLAIF, verifiable math and code samples, and the diversity controls that stop collapse
datatrainingalignment
07
3 min
Quantization
Trading precision for memory: what NF4, GPTQ and AWQ each cost in quality, when GGUF beats EXL2, and why KV cache quantization matters most
quantizationinferencecost