Course Format and Delivery
Delivery Method: LiveOnline
Schedule: 3 hours
All sessions are delivered live by an expert instructor in a fully interactive online environment.
*20% off for group bookings when booking 3 or more attendees from the same organization on the same course dates in the same transaction.
About this course
Pre-trained models are powerful generalists, but most enterprise use cases demand domain specific behavior: internal terminology, company style guides, proprietary classification schemas, or regulatory language. Fine-tuning bridges that gap. LoRA (Low-Rank Adaptation) reduces trainable parameters to 0.1-1% of the total while recovering 90-95% of full fine-tuning quality. QLoRA adds 4-bit quantization, cutting memory requirements by ~75%. Together they make fine-tuning accessible on modest hardware.
This half-day workshop teaches the complete fine-tuning workflow — from dataset preparation through training to systematic evaluation — using small models (distilgpt2 124M, TinyLlama 1.1B, Phi-3-mini 3.8B) that train in minutes on CPU-only Codespace environments with 4 vCPUs. Participants experience the full lifecycle: prepare a Hugging Face dataset, run a baseline evaluation, fine-tune with LoRA, fine-tune with QLoRA, compare results, and use evaluation frameworks (lm-evaluation-harness, custom metrics) to measure improvement quantitatively.
What you will need
Basic Python proficiency and conceptual understanding of what LLMs are. No prior ML training experience required. No GPU required — all labs run on standard 4-vCPU GitHub Codespaces.
Topics Covered
When to fine-tune vs. prompt-engineer vs. RAG, dataset curation and formatting (instruction/completion pairs, chat templates, Hugging Face datasets), the Hugging Face Trainer and TRL (Transformer Reinforcement Learning) library, full fine-tuning on a tiny model (distilgpt2) to establish baseline understanding, LoRA theory and practice (rank, alpha, target modules, DoRA), QLoRA (4-bit NormalFloat quantization + LoRA adapters via bitsandbytes), adapter merging and export, and systematic evaluation using EleutherAI lm-evaluation-harness, task-specific metrics, and before/after comparison methodology. Labs use progressively larger models to show how LoRA/QLoRA scale.
