Cover of Fine-Tuning and Adapting Open LLMs
DRM-free · Yours to keep forever
AI & LLMs

Fine-Tuning and Adapting Open LLMs

LoRA, quantization and instruction tuning on your own data

By Houssam Kodad

PDF 216 pages Advanced English

One-time purchase

€28.95

VAT included
where applicable

Download sample
  • Instant download after purchase
  • Readable on any device
  • Free updates to this edition
  • Secure checkout

About this book

What's inside

When prompting hits its limits, fine-tuning an open model on your own data is the next move — but it's full of expensive ways to waste a week. This book gives you a grounded process for adapting open LLMs with LoRA and QLoRA, from building a clean instruction dataset to evaluating and serving the result. You'll learn what fine-tuning can and can't fix, and how to do it on a realistic GPU budget.

What you'll learn

Skills you'll walk away with

  • Decide when to fine-tune versus prompt or use RAG
  • Build and clean an instruction-tuning dataset
  • Apply LoRA and QLoRA for parameter-efficient tuning
  • Use quantization to train and serve on modest GPUs
  • Set hyperparameters that converge without overfitting
  • Evaluate a tuned model against honest baselines
  • Serve adapters efficiently in production

Table of contents

9 chapters
  1. 01

    When Fine-Tuning Is the Right Tool

    • · Prompting vs RAG vs tuning
    • · What tuning can and cannot fix
    • · Cost and effort reality check
  2. 02

    Datasets Make or Break It

    • · Instruction data design
    • · Cleaning and deduplication
    • · Synthetic data with care
  3. 03

    Parameter-Efficient Fine-Tuning

    • · Full vs LoRA tuning
    • · How LoRA adapters work
    • · Targeting the right layers
  4. 04

    Quantization for Modest GPUs

    • · 8-bit and 4-bit basics
    • · QLoRA end to end
    • · Memory and throughput trade-offs
  5. 05

    Running a Training Job

    • · Hyperparameters that matter
    • · Monitoring loss and stability
    • · Checkpointing and resuming
  6. 06

    Alignment and Preference Tuning

    • · Supervised fine-tuning
    • · DPO and preference data
    • · Avoiding capability regressions
  7. 07

    Evaluating a Tuned Model

    • · Task-specific benchmarks
    • · Regression against the base model
    • · Human evaluation
  8. 08

    Serving Fine-Tuned Models

    • · Merging vs serving adapters
    • · Multi-adapter serving
    • · Latency and batching
  9. 09

    Maintaining Adapted Models

    • · Versioning data and weights
    • · Re-tuning on new data
    • · Governance and licensing

This is the full chapter list — exactly what you'll receive in the PDF.