🔧
LLM Fine-Tuning and Customization
An L200 intermediate, hands-on course. You know what an LLM is — now learn to customize it. This practical course walks you step by step through fine-tuning an open-source model (Llama, Mistral) on your own data with LoRA/QLoRA. From dataset preparation to putting the model into service, you will master the full customization pipeline.
Duration
3 days
Level
Intermediate
Price
9.99 EUR/month (all courses included)
Max group
12 participants
What you will learn
+Choose between fine-tuning, RAG, and prompt engineering based on the use case
+Prepare a quality dataset in standard formats (JSONL, Alpaca, ShareGPT)
+Fine-tune a model with LoRA and QLoRA on a consumer GPU
+Evaluate a fine-tuned model with quantitative and qualitative metrics
+Deploy a customized model with vLLM, TGI, or Ollama
+Optimize costs: base-model choice, post-training quantization
Course program
Module 1: When to fine-tune: a decision tree
3h30- Fine-tuning vs RAG vs prompt engineering: decision matrix
- Types of fine-tuning: full, LoRA, QLoRA, prefix-tuning
- Knowledge distillation: train a small 'student' model from a large 'teacher' to cut serving costs
- Hardware and budget prerequisites: GPU VRAM, compute time
- Workshop: evaluate 5 use cases and recommend the right approach
Module 2: Data preparation
3h30- Dataset formats: JSONL, Alpaca, ShareGPT, conversation
- Cleaning and validation: deduplication, inconsistencies, bias detection
- Quality metrics: diversity, coverage, distribution
- Workshop: build a 500-example dataset from raw data
Module 3: Fine-tuning with LoRA and QLoRA
3h30- LoRA: theory (low-rank matrices) and implementation with PEFT
- QLoRA: combining quantization and LoRA for consumer GPUs
- Tools: Hugging Face Transformers, Axolotl, Unsloth
- Workshop: fine-tune Mistral 7B on a business dataset
Module 4: Evaluation and benchmarking
3h30- Automatic metrics: perplexity, BLEU, ROUGE
- LLM-as-judge: using Claude/GPT to evaluate responses
- Human evaluation: scoring rubrics, inter-annotator agreement
- Workshop: build a complete evaluation suite for your model
Module 5: Deployment and serving
3h30- vLLM: high-performance serving with paged attention
- Text Generation Inference (TGI): the Hugging Face standard
- Ollama: deploy locally for prototyping and light production
- Workshop: deploy your fine-tuned model and benchmark latencies
Module 6: Capstone project: client case
3h30- Full case study: fine-tune a customer-support model
- End-to-end pipeline: raw data → dataset → fine-tuning → evaluation → deployment
- Cost optimization: base-model choice, post-training quantization
- Presentation and trainer feedback
Ready to get started?
9.99 EUR/month — All courses included, cancel anytime
Go further
Recommended video resources
A curated selection of videos from leading experts to deepen your understanding of each course module.
Module 1
Module 2
Module 3
Module 5
ⓘ These videos are external content produced by independent creators and are not owned by Talki Academy. They are recommended for educational purposes to complement and illustrate the course material.