Model Adaptation · 3 to 4 hours
Design a model specialization program
Design a specialization program for a fictional hint model within GPU-hour and expert-review budgets: choose learning signals and parameterization, protect splits and evaluation, and plan a release with rollback. Then investigate a candidate that won a preference comparison while experts found wrong explanations.
System design and experimental reasoning
4 h
Used in interview rounds:
AI System Design (Hard)Overview
Design an adaptation program that stays trustworthy across model changes, preference data, domain shifts, and releases. The hard part is choosing which behavior to optimize while preserving other capabilities and keeping reproducible evidence: a more sophisticated optimizer cannot repair an ambiguous objective or a contaminated evaluation. You practice on a fictional hint model with fixed GPU-hour and expert-review budgets, then investigate a candidate that won a preference comparison while experts reported wrong explanations.
Study time: 3 to 4 hours
Interview format: System design discussion with budget arithmetic and a failure investigation
By the end you can
- Separate the learning signal, the trainable parameters, the numeric representation, and the deployment form, and justify each choice from a measured behavior gap
- Compute LoRA trainable entries and adapter size for a stated configuration, and explain what rank does and does not limit
- Explain what the DPO objective optimizes relative to a reference policy, and how preference labels can reward artifacts instead of quality
- Plan training runs, repeated seeds, and evaluation within a GPU-hour and expert-review budget
- Trace an experiment result that conflicts with expert review to training targets, labels, pair generation, or evaluation coverage before choosing a remedy
- Specify a release manifest, per-slice gates, rollback, and data-removal handling for an adapted model
Prerequisites
Helpful background: Familiarity with gradients, probability, and experimental design helps.
This topic covers choosing between prompting, retrieval, and fine-tuning, validating training data, and proving that specialization helped. Training infrastructure and distributed training are out of scope.
In this guide
- Separate objectives from parameterization
- Low-rank adaptation mechanics
- Quantized adaptation and export
- Continued pretraining and SFT
- Preference optimization
- Reward-based adaptation
- Distillation and synthetic supervision
- Multi-task and tenant specialization
- Causal evidence and ablations
- Deployment and release architecture
- Data correction and removal
- Scenario: specialize a hint model on a fixed budget
- Failure investigation: last quarter's candidate
Hard questions are part of Premium: the study guide, reference answer, rubric, and exercises unlock with Premium.
Related on CodeSnatch
Lessons, problems, and articles that build the skills this question uses.
System Design
A model registry with data lineage, shadow traffic, canary promotion, and one-step rollback maps onto this release plan.
AI Engineering
Practice a ship decision when offline ratings and later outcomes disagree, with intervals, a rollback trigger, and reversal evidence.
Behavioral Interview
Frames stories on experiment rigor and consent boundaries for training data, signals ML design rounds also grade.
System Design
Traces an erasure request through warehouses, backups, logs, and caches, a pattern for removal across derived copies.
Interview experience
A first-person account of deriving the logistic loss gradient by hand, the sigmoid form that the DPO loss builds on.
Practice the interview question
Answer it in your own words before you look at the reference answer.