Week 3: Fine-Tuning and RAG

• Designing instruction-response datasets
• Cleaning noisy examples and removing leakage
• Balancing diversity and style consistency

• Parameter-efficient fine-tuning with minimal GPU cost
• Choosing rank, alpha, and target modules
• Checkpoint management and rollback strategy

# Pseudo setup for PEFT config
lora_r = 16
lora_alpha = 32
dropout = 0.05
batch_size = 4
grad_accum = 8

• Chunking, embeddings, and vector search strategies
• Hybrid retrieval: keyword + semantic reranking
• Citation-first generation and confidence scoring

• Task-specific eval sets and golden responses
• Faithfulness, groundedness, and answer completeness
• Regression testing before deployment

• Lab 10: Prepare and validate domain dataset
• Lab 11: Fine-tune with LoRA
• Lab 12: Create embedding pipeline and retriever
• Lab 13: Build RAG QA API with citations
• Lab 14: Evaluate and iterate with error-driven fixes

When to use Fine-Tuning vs RAG (simple rule)

• Use Fine-Tuning when you want a model to follow your style, tone, or task format consistently.
• Use RAG when answers must use latest documents, policies, manuals, or private company knowledge.
• Use both when you need domain style + factual grounding from fresh documents.

Daily action plan (2 to 3 hours/day)

1. Day 1: Build a small, clean dataset of 100 to 300 high-quality instruction-answer pairs.
2. Day 2: Run a basic LoRA fine-tune. Compare base model and tuned model on the same 20 test prompts.
3. Day 3: Build a retrieval pipeline: chunk docs, create embeddings, and query top relevant chunks.
4. Day 4: Add citation output and confidence checks. Reject answers when evidence is weak.
5. Day 5: Integrate tuned style + RAG grounding into one knowledge copilot.

Quality checklist before moving to Week 4

• Fine-tuned model response style is visibly more aligned with your target format
• RAG answers include source citation for each key claim
• At least 20-question evaluation set with measured scores
• Failure cases are documented with planned fixes

Assignment to complete Week 3

Build a Knowledge Copilot for one domain (for example: HR policy, product docs, SOP manuals, or legal templates). The assistant must answer with citations and decline when confidence is low.

Scoring rubric