← All courses

Course 8 · Expert / Deep-Dive

Custom Model Engineering — Fine-Tuning & Quantization

16 hours · 4 sessions × 4 hours · Max 10 participants

A reproducible post-training pipeline: locked dataset → QLoRA → checkpoint → adapter → merged model → benchmark → GGUF → quantized Ollama deployment.

Who should attend

Highly advanced ML engineers, technical founders, and well-funded development departments.

Where and how

  • In person in Haiphong, Vietnam.
  • English-language delivery.
  • Bring your own laptop and build as you learn.
  • Open groups and private B2B cohorts available.

What we cover

Session 1 — Dataset Architecture — Decide between fine-tuning and RAG. Build instruction datasets and chat templates. Run token diagnostics, cleaning, and versioned private datasets.

Session 2 — Cloud Compute & PEFT — Provision GPUs, configure SSH and persistence, learn LoRA and QLoRA mechanics, set hyperparameters, run smoke tests, and recover checkpoints.

Session 3 — Training & Telemetry — Run resilient training with loss and VRAM telemetry. Control experiment forks, export adapters, and complete a clean model merge.

Session 4 — Evaluation & Porting — Use a frozen benchmark. Convert to GGUF, apply Q8 and Q4 quantization, integrate Ollama, and complete the technical handover.

You will learn to

  • Decide when fine-tuning is justified.
  • Engineer and version template-correct datasets.
  • Provision and verify cloud GPU environments.
  • Configure LoRA and QLoRA for recovery-safe training.
  • Interpret loss and telemetry with experiment discipline.
  • Merge and evaluate model artifacts.
  • Convert to GGUF, quantize, and serve locally with Ollama.

Investment

B2C
15,000,000 VND / person
B2B private cohort
70,000,000–110,000,000 VND / private cohort

Recommended range. GPU/cloud compute and proprietary data preparation are quoted separately.

Completion

Attend, complete the assessment, and finish the capstone. We will give you a course completion certificate.

For private cohorts, we can adapt the examples to a cleaned-up company use case after a short technical call.