The Case for Specialized Models
Follow a bookshop request, learn what makes a model a specialist, and see where prompting and fine-tuning fit.
Fine-Tuning Small Language Models
Build and evaluate a router, invoice extractor, and SQL specialist with LoRA, QLoRA, synthetic data, model scaling, and a measured path to deployment.
Download as an EPUB 21 chapters — read offline on any e-reader
-
-
Do You Actually Need Fine-Tuning?
Compare instructions, demonstrations, dynamic example selection, retrieval, and training by following the evidence a bookshop request needs.
-
What Fine-Tuning Actually Changes
Follow one small weight update, connect it to the router's training examples, and understand what full fine-tuning, LoRA, and QLoRA change and save.
-
Building the Fine-Tuning Laboratory
Prepare the router project, check one answer, and save a reproducible baseline before looking inside the inference code.
-
Establishing the Baseline
Read the prompting results, investigate missed routes, and build up the quality and timing measures that a fine-tuned router must improve.
-
Building Training Data
Turn a routing policy into supervised examples, separate evaluation before expansion, and distinguish more rows from more useful coverage.
-
Your First Full Fine-Tune
Train every parameter of the small router, supervise only destination tokens, and judge the saved checkpoint against an unchanged model.
-
LoRA: Train Less, Learn More
Understand low-rank updates, train a small router adapter, and compare its behavior, memory, and artifacts with full fine-tuning.
-
QLoRA: Fine-Tuning Under Memory Constraints
Keep the base in four-bit storage, train higher-precision adapters, and measure the quality and runtime costs of that memory trade.
-
How Much Training Data Do You Need?
Measure six dataset sizes, compare useful coverage with repeated rows, and interpret a learning curve without confusing data volume with evidence.
-
Synthetic Data Generation
Use a larger model to propose training text, preserve known answer fields, filter candidates, and compare generated data without inventing human supervision.
-
Failure-Driven Fine-Tuning
Turn observed errors into a targeted repair set, keep a matched training control, and use a fresh audit without recycling the test as a source of labels.
-
From Classification to Structured Generation
Turn bookshop invoice text into a strict record, compare prompting with LoRA, and separate valid JSON from correct facts.
-
When the Real World Gets Messy
Stress the invoice extractor with missing facts, conflicting totals, ambiguous dates, noise, and new layouts, then measure what targeted augmentation repairs.
-
Fine-Tuning for a Harder Generative Task
Train a bookshop SQL specialist and grade query results against independent oracles, with explicit limits on syntax, permissions, and semantic correctness.
-
When the Small Model Hits Its Limit
Compare Qwen3 0.6B, 1.7B, 4B, and 8B under one SQL training plan, then choose capacity from task quality and measured costs.
-
Prompt, Retrieve, Fine-Tune, or Scale?
Use the book's measured router, extractor, and SQL results to choose an intervention from the failure it can actually repair.
-
Building the Multi-Model Specialist System
Connect the router, extractor, and SQL specialist through typed inputs, independent authorization, validation, and an explicit fallback path.
-
The Economics of Small Models
Compare API calls, hosted large models, and specialist pipelines with explicit utilization, fallback, staffing, and amortization assumptions.
-
Production Fine-Tuning
Turn a trained adapter into a versioned release, gate changes with regressions, monitor drift, and rehearse rollback without confusing a local demo with production readiness.
-
Appendix A: Router Laboratory Reference
Dependency roles, inference calls, example-selection rules, and full reports for the prompting baseline in Chapters 4–5.