Module 5 of 5 · 60 min
Model Customization Capstone: Train, Align & Serve
End-to-end capstone: Curate instruction dataset, execute QLoRA training, align with DPO, evaluate benchmarks, and deploy an AWQ quantized endpoint.
Core concept
By the end
You will be able to
- Deliver an end-to-end custom fine-tuned model pipeline meeting accuracy and latency targets.
- Execute QLoRA fine-tuning and DPO alignment.
- Quantize to AWQ and verify serving endpoint performance.
01
Custom Model Capstone Requirements
In this capstone, you will demonstrate full lifecycle model customization: from raw domain data curation and QLoRA fine-tuning, through DPO alignment, automated benchmark validation, and final quantized serving.
Practice activity
Deliver Complete Fine-Tuned Domain Model
- Curate 250 domain instruction pairs.
- Execute QLoRA adapter training and DPO alignment.
- Export to AWQ format and verify vLLM serving latency and benchmark accuracy.
What to produce
- Training script, loss logs, DPO win-rate evaluation report, and endpoint test receipts.
Reflect before continuing
How do you safeguard against domain catastrophic forgetting when tailoring specialized models?
Evidence
Sources and verification
- An Overview of Catastrophic Forgetting in Deep LearningArXiv · verified 2026-08-22
Knowledge check
Make it stick.
Choose the strongest answer for each question. Your attempts become part of your account transcript.