You've covered the full pipeline from theory to deployment. Now apply it. This capstone presents three real challenges to test your skills — then you earn your certificate.
When prompts fail. The spectrum from prompting to pre-training. When fine-tuning wins and loses.
The three formats. Five quality filters. JSONL output for training frameworks.
Low-rank adaptation. Rank and alpha. QLoRA on consumer hardware.
SFT, reward models, PPO loop. DPO as the simpler modern alternative.
Four-layer eval stack. Catastrophic forgetting detection. LLM-as-judge.
Merge LoRA, quantize with GPTQ or GGUF, serve via FastAPI or HF Endpoints.
Choose a domain you know: your industry, your job, or a topic you're deeply familiar with. Collect or write 100 instruction-response pairs. Apply all five quality filters from Module 2. Format them as JSONL. Write a short paragraph explaining: what task does this dataset teach, and what format consistency rule did you enforce?
Using the code from Module 3, configure a LoRA fine-tuning run for your dataset. You don't need to actually run it (that requires GPU resources). Instead: write out the LoraConfig with your chosen rank and alpha, explain why you chose those values, and list which target modules you would use and why. Then write the evaluation plan from Module 5: what metrics would you use, and how would you check for catastrophic forgetting?
You've fine-tuned a 7B model for internal customer support at a company handling 300 support tickets per day. Average response length is 150 tokens. The team needs responses in under 2 seconds. Internal security policy does not allow data to leave the company's servers. Write a deployment plan: Which quantization format? Which serving approach? Why? What would the estimated infrastructure look like?
Enter your name below. The certificate generates automatically. Click to edit any field.
Book a free 30-minute call to discuss your fine-tuning use case and what it would take to get a pilot running.
Book with Arjun