Chapter Summary

Chapter Summary

Key Points

• 1.

  LoRA fine-tuning is the practical default. It modifies only 0.1% of parameters with rank-$r$ updates $\Delta\mathbf{W} = \mathbf{B}\mathbf{A}$, achieving near-full fine-tuning quality. QLoRA further reduces memory via 4-bit quantization of base weights.

• 2.

  Data quality beats quantity for fine-tuning. 1000-5000 carefully curated instruction-response pairs suffice for most domain adaptation tasks. Always validate on a held-out set.

• 3.

  Training from scratch teaches the full pipeline. nanoGPT-style implementations demystify LLMs: data preparation, model architecture, training loop with gradient clipping and cosine LR.

• 4.

  Instruction tuning creates chat models. Compute loss only on response tokens, use proper chat templates, and evaluate on diverse benchmarks.

• 5.

  Multimodal models combine vision and language. Vision encoders project image features into the LLM's embedding space for joint reasoning.