Module 16: Post-Training, Adaptation, And Evaluations

Introduction

A broad model may be capable and still be a poor fit for a product. It may return the wrong format, invent details, mishandle exceptions, or take too much time and money. Post-training changes learned behavior after pretraining. Evaluation helps establish what improved, under what conditions, and what may have become worse.

We will connect supervised fine-tuning, preference optimization, reinforcement learning, parameter-efficient adaptation, distillation, and compression to evaluation. Each method addresses particular problems. Your project is a twenty-five-case evaluation packet and a model-selection memo grounded in the results.

Review Modules 1, 13, and 14 for training, loss, gradients, data splits, parameters, context, and generation. Study adaptation first, evaluation design next, then the complete case packet. Keep an unchanged baseline. Without it, it is much harder to know whether an intervention helped.

Planned lessons

Diagnose The Failure Before Choosing Training

Supervised Fine-Tuning, Or SFT

Preference Data And The Reward-Model Approach

Direct Preference Optimization, Or DPO

Reinforcement Learning And Verifiable Rewards

Parameter-Efficient Fine-Tuning And LoRA

Catastrophic Forgetting, Regressions, And Retained Capability

Distillation, Pruning, Quantization, And Speculative Decoding

Define “better” As A Set Of Task Conditions

Case Design, Slices, And Coverage

Rubrics, References, And Human Review

Small Samples And Honest Uncertainty

Latency, Cost, Quality, And Successful Work

Red Teaming, Adversarial Tests, And Controlled Recovery

Offline Evaluation, Shadowing, And Production Experiments

A Model-Selection Memo That Supports A Decision