LLM Engineering
LLM Adaptation and Runtime
Data, Post-Training, Inference, and Model Change
Komal Nakrani
First edition/LLM Engineering · Volume 2
LLM Adaptation and Runtime
Data, Post-Training, Inference, and Model Change
A professional guide to evidence-led model adaptation, data design, post-training, runtime qualification, and reversible change.
- Edition
- First edition
- Version
- 1.0.0
- Published
Table of contents
01Start From a Frozen Behavior BaselineTurn a negative adaptation referral into a falsifiable investigation anchored to an exact, replayable behavior baseline.95 min02Inspect the Model and Tokenizer BoundaryInspect model, tokenizer, template, precision, and runtime as separate compatibility surfaces without turning architecture into a quality forecast.95 min03Choose the Smallest Adaptation LadderMap each residual failure to the smallest plausible intervention, reject mismatched methods, and attach disconfirmation and stop rules before training.100 min04Specify the Data RecipeTurn a conditional adaptation hypothesis into a reproducible, authority-aware data recipe before collecting or transforming examples.100 min05Curate, Deduplicate, and SeparateExecute the data recipe as an auditable pipeline that preserves rejected records, detects indirect overlap, and freezes purpose-separated dataset vaults.105 min06Build Instruction and Demonstration DataTurn clean synthetic partitions into evidence-linked, template-correct instruction records with intentional loss regions and visible coverage gaps.105 min07Build Preference and Feedback DataTurn comparative judgments into criterion-specific, randomized, bias-audited evidence without treating preference as truth.105 min08Preserve Retention and Control CasesFreeze target, retention, multilingual, abstention, safety, and no-change lanes so adaptation cannot hide forbidden regressions.105 min09Establish the Training ExperimentSpecify an inspectable, resumable training experiment whose state, resources, behavioral hooks, and stop rules are frozen before execution.110 min10Supervise the ModelSeparate supervised token optimization from held-out behavior evidence, then select or reject checkpoints across target, retention, language, control, and resource slices.110 min11Adapt Efficiently With PEFTEvaluate low-rank adapters as traceable behavior-and-resource candidates whose base, tokenizer, template, runtime, quantization, and merge state must remain compatible.110 min12Optimize Preferences CarefullyTreat preference optimization as a bounded proxy experiment whose value must survive independent target, retention, language, control, and bias evaluation.115 min13Decide on Distillation or Continued PretrainingChoose advanced adaptation only when its mechanism, transfer source, authority, and disconfirming evidence match a distinct residual gap.110 min14Package and Version the Model SystemBind every behavior-facing component, lineage link, compatibility result, limitation, and rollback state into one fail-closed release identity.110 min15Reason About Inference Memory and ThroughputBuild a workload-specific resource model that separates weights, KV state, prefill, decode, concurrency, throughput, and latency tails.115 min16Compress, Serve, and Preserve BehaviorSelect serving variants on an expiring behavior-resource frontier, with exact compatibility, protected replay, and graceful degradation.115 min17Release, Diagnose, and Evolve Adapted ModelsClose the series with an authority-preserving release, incident, rollback, requalification, and durable-learning loop across the complete model-system dossier.120 min