Appendix A: Back Matter

Key Takeaways

• The glossary fixes the vocabulary for behavior, knowledge, adapters, splits, and release gates.
• The checklist turns the book into an implementation sequence from diagnosis through operation.
• The source register keeps external citations explicit and avoids invented authority.

Glossary

Adapter: A small set of trainable parameters (as in LoRA) added on top of a frozen base model. Detachable, composable, and the basis for instant rollback.

Baseline: The measured performance of the cheaper alternatives (zero-shot, prompted, few-shot, retrieval) on your task, established before training so the fine-tune has a meaningful number to beat.

Behavior: How a model responds: tone, format, task steps, tool discipline, domain phrasing. Lives in the weights; the right thing to fine-tune. Contrast knowledge.

Catastrophic forgetting: General capability lost as a side effect of fine-tuning, because the optimizer moves weights that encoded unrelated skills. Caught by regression evals.

Contamination: Overlap between training and test data (exact, near-duplicate, entity, temporal, or pretraining) that inflates the eval and makes a model look good that isn't.

Correction: A captured production failure plus its fix; the highest-value demonstration, feeding both the training set and the regression suite (the flywheel).

Data card / Datasheet: A document recording a dataset's sources, owners, licenses, PII/consent status, labeler-instruction version, measured quality, and splits. Part of lineage.

Demonstration: An input paired with its correct, complete output. The signal for supervised fine-tuning.

Distillation: Training a small student model to reproduce a large teacher model's behavior. Inherits the teacher's mistakes; evaluate against ground truth, not teacher agreement.

DPO (Direct Preference Optimization): Preference tuning without a separate reward model or RL loop; an SFT-like loss that raises the likelihood of preferred over dispreferred responses.

Drift: Quality degradation over time with no change to the weights, because the world the training data described moved on. Detected by monitoring; triggers retraining.

Generalist tax: The capability a frontier model offers but a narrow task never uses, billed on every call. The thing small-model specialization attacks.

Golden test set: A frozen, versioned, meticulously-correct set of test cases used to compare model versions on a stable ruler. Never trained on.

Knowledge: What a model treats as true: current facts, business state, policy. Lives in sources, not weights; retrieve or fetch it, never freeze it. Contrast behavior.

Lineage: The traceable chain from a deployed model back to the exact data, config, and run that produced it (dataset card + model card + run manifest). Required for reproduction, rollback, deletion, and audit.

LoRA (Low-Rank Adaptation): Fine-tuning that freezes the base and trains small low-rank adapters. Cheap, forgetting-bounded, perfectly reversible by detaching the adapter.

Model card: A document recording what a model is: base, method, eval results across all axes, intended use, limitations, and rollback target.

Preference: A judgment that one response is better than another for the same input. The signal for DPO/preference tuning, for properties with no single right answer.

QLoRA: LoRA on a 4-bit-quantized base, cutting memory enough to fine-tune large models on a single GPU with adapters kept at full precision.

Regression wall: The release gate that requires the new model to improve the target task while not regressing on protected axes (safety, format, key slices). A regression on any protected axis blocks the release.

Release gate: An automated set of preconditions (clean splits, beat baseline by a margin, regression wall, validated judge) a fine-tune must clear before shadow and canary rollout.

SFT (Supervised Fine-Tuning): The training objective of learning from labeled demonstrations. Orthogonal to the parameterization (full / LoRA / QLoRA).

Slot pattern: Writing training examples with placeholders for current facts so the model learns the phrasing (behavior) and the fact is filled by retrieval at inference time.

Stable repeated task: A high-volume task whose desired output is a structural shape (not a fact) that rarely changes. The canonical fine-tuning case.

Synthetic data: Examples generated by a model. Powerful for coverage and labeling cost; dangerous via inherited error, distribution narrowing, and recursive collapse. A candidate until verified.

TRAIN: The book's decision framework, and the fine-tuning companion to Human in the Loop Is Not a Plan and A Field Guide to Evals:Task stability, Required knowledge, Available examples, Impact of mistakes, Necessary evaluation. All five must pass before training.

---

Implementation Checklist

A team's fine-tuning decision and execution is sound when it can answer yes, with evidence to each of these. Grouped by movement.

Diagnosis (Movement I)

• The symptom was named and sorted into the taxonomy (behavior vs. knowledge vs. state vs. permission) before "fine-tune" was written on a ticket.
• No current fact, per-request state, permission, or fast-changing policy is a fine-tune target; each routes to retrieval, tools, or policy-in-code.
• The five false diagnoses were ruled out: the distinguishing test was run for "doesn't know," "needs our data," "too verbose," "fails sometimes," and "must be cheaper."
• "It fails sometimes" was converted into clustered, characterized failures before any training decision.
• The lightest adequate tool on the ladder was tried and shown to fall short before climbing to fine-tuning.

The case for training (Movement II)

• The target is a stable, repeated, pattern-shaped task, green on the task-stability scorecard, not red on stability or task-ness.
• Style/phrasing/tool-discipline fine-tunes slot out facts and raise, not lower, the grounding bar for the knowledge twin.
• Small-model specialization is justified by a break-even calculation including data curation and ongoing ops, not just GPU cost.

Data (Movement III)

• The signal kind (demonstration / correction / preference) matches the problem (right answer vs. better answer).
• Inter-annotator agreement was measured; contested items were reviewed; labeler instructions are versioned.
• Coverage includes edge cases, negatives, and refusals, not just the easy middle.
• The validator passes: no contradictions (release blocker), no invalid labels, no severe imbalance, negatives present.
• Splits are by entity and temporal where relevant; the contamination report is clean across every split boundary; a frozen golden set exists.
• Synthetic data is verified by a separate checker, diversity-capped, blended with real data, and provenance-tagged.

Methods (Movement IV)

• The objective (SFT/DPO/distillation) and the parameterization (LoRA/QLoRA/full) were chosen as separate, justified decisions.
• LoRA/QLoRA is the default; full fine-tuning is used only where an eval proves the adapter can't reach the bar.
• The run pins the base, dataset hash, and seed; training stops on the validation metric, not the training loss.
• The right component was chosen, retriever or router fine-tuned where the failure actually lives, not the generator by reflex.

Evaluation and operations (Movements V-VI)

• A baseline was measured before training; the fine-tune must beat the best cheaper combination by a margin beyond eval noise.
• The eval suite covers task, regression, format, safety/truthfulness, and slices; the regression wall blocks on any protected-axis regression.
• Any LLM judge is validated against humans, uses a different family, and randomizes order.
• The release gate is automated; passing it earns shadow + canary with auto-rollback, not a full launch.
• Lineage exists (dataset card + model card + run manifest); the rollback target is a real, exact previous state.
• Drift monitoring watches input, output, quality, and protected-slice drift; crossing a threshold triggers retraining via the correction flywheel.
• Deletable PII and fast-changing policy were kept out of the weights; the dataset (not the trained model) is treated as the portable asset.
• An incident runbook exists with an instant adapter-detach rollback as step two.

---

Research and Source Register

Sources grouped by chapter. A source appears under a chapter only if that chapter actually uses it to support a claim.

Front matter / Introduction: synthetic; draws on the book's own argument. No external citations.

Ch. 1, The Support Bot That Knew the Old Product

• OpenAI: Model optimization guide
• OpenAI: Supervised fine-tuning
• OpenAI: Fine-tuning best practices
• Retrieval-Augmented Generation (Lewis et al.)
• Training language models to follow instructions (InstructGPT)

Ch. 2, What Fine-Tuning Actually Changes

• OpenAI: Supervised fine-tuning
• Training language models to follow instructions (InstructGPT)
• LoRA: Low-Rank Adaptation of Large Language Models
• An Empirical Study of Catastrophic Forgetting in LLMs during Continual Fine-tuning
• Physics of Language Models: Knowledge Capacity Scaling Laws

Ch. 3, Five False Diagnoses

• Retrieval-Augmented Generation (Lewis et al.)
• OpenAI: Model optimization guide
• OpenAI: Fine-tuning best practices
• Lost in the Middle: How Language Models Use Long Contexts
• Chain-of-Thought Prompting Elicits Reasoning in LLMs

Ch. 4, The Customization Menu and a Decision Tree

• OpenAI: Model optimization guide
• Retrieval-Augmented Generation (Lewis et al.)
• Chain-of-Thought Prompting Elicits Reasoning in LLMs
• Toolformer: Language Models Can Teach Themselves to Use Tools
• LoRA: Low-Rank Adaptation of Large Language Models

Ch. 5, Format, Behavior, and the Shape of a Repeated Task

• OpenAI: Supervised fine-tuning
• OpenAI: Fine-tuning best practices
• Stanford Alpaca
• Self-Instruct
• Training language models to follow instructions (InstructGPT)

Ch. 6, House Style, Domain Phrasing, and Tool Discipline

• Training language models to follow instructions (InstructGPT)
• Toolformer: Language Models Can Teach Themselves to Use Tools
• OpenAI: Supervised fine-tuning
• Self-Instruct
• OpenAI: Function calling guide

Ch. 7, Specializing Small Models and Distilling Down

• Distilling the Knowledge in a Neural Network (Hinton et al.)
• Textbooks Are All You Need (phi-1)
• Phi-3 Technical Report
• QLoRA: Efficient Finetuning of Quantized LLMs
• OpenAI: Model optimization guide

Ch. 8, Demonstrations, Corrections, and Preferences

• Training language models to follow instructions (InstructGPT)
• Direct Preference Optimization
• OpenAI: Fine-tuning best practices
• Constitutional AI: Harmlessness from AI Feedback
• Self-Instruct

Ch. 9, Labels, Disagreement, and Coverage

• OpenAI: Fine-tuning best practices
• Training language models to follow instructions (InstructGPT)
• Data Quality for Machine Learning Tasks (KDD)
• Datasheets for Datasets
• Pervasive Label Errors in Test Sets (Northcutt et al.)

Ch. 10, Contamination, Leakage, and the Splits That Save You

• Generalization or Memorization: Data Contamination and Trustworthy Evaluation for LLMs
• OpenAI: Fine-tuning best practices
• QLoRA: Efficient Finetuning of Quantized LLMs
• Pervasive Label Errors in Test Sets (Northcutt et al.)
• Documenting Large Webtext Corpora: A Case Study on C4

Ch. 11: Synthetic Data: When It Helps, When It Poisons

• Self-Instruct
• Stanford Alpaca
• Textbooks Are All You Need (phi-1)
• AI models collapse when trained on recursively generated data (Nature)
• Constitutional AI: Harmlessness from AI Feedback

Ch. 12, SFT, LoRA, and QLoRA in Practical Terms

• LoRA: Low-Rank Adaptation of Large Language Models
• QLoRA: Efficient Finetuning of Quantized LLMs
• Hugging Face PEFT documentation
• OpenAI: Supervised fine-tuning
• An Empirical Study of Catastrophic Forgetting in LLMs during Continual Fine-tuning

Ch. 13, Preference Tuning, DPO, and Distillation

• Direct Preference Optimization
• Training language models to follow instructions (InstructGPT)
• Constitutional AI: Harmlessness from AI Feedback
• Distilling the Knowledge in a Neural Network (Hinton et al.)
• QLoRA: Efficient Finetuning of Quantized LLMs

Ch. 14, What to Fine-Tune: Generators, Retrievers, and Routers

• Dense Passage Retrieval for Open-Domain Question Answering
• Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
• Retrieval-Augmented Generation (Lewis et al.)
• LoRA: Low-Rank Adaptation of Large Language Models
• QLoRA: Efficient Finetuning of Quantized LLMs

Ch. 15, Baselines, Regression Walls, and the Release Gate

• OpenAI: Evals guide
• OpenAI Evals repository
• RAGAS: Automated Evaluation of Retrieval Augmented Generation
• TruthfulQA: Measuring How Models Mimic Human Falsehoods
• QLoRA: Efficient Finetuning of Quantized LLMs

Ch. 16, Versioning, Lineage, Drift, and Retirement

• NIST AI Risk Management Framework
• GDPR Article 17: Right to erasure
• Model Cards for Model Reporting
• Datasheets for Datasets
• OpenAI: Supervised fine-tuning

Ch. 17, Ten Playbooks for the Decision Meeting

• OpenAI: Fine-tuning best practices
• Retrieval-Augmented Generation (Lewis et al.)
• Direct Preference Optimization
• Toolformer: Language Models Can Teach Themselves to Use Tools
• NIST AI Risk Management Framework

---