Train and customise your own Large Language Models. Master LoRA, QLoRA, PEFT techniques, RLHF, and DPO alignment -- then deploy your fine-tuned models at scale with vLLM and TGI.
Master the Hugging Face platform: Transformers library, model hub, tokenizers, datasets, and the Trainer API. Your complete toolkit for working with LLMs.
Parameter-efficient fine-tuning with Low-Rank Adaptation. Understand the math behind LoRA, implement QLoRA for memory-efficient training, and configure rank, alpha, and target modules.
Beyond LoRA: prefix tuning, prompt tuning, IA3, and adapter methods. Compare approaches, understand trade-offs, and choose the right PEFT technique for your use case.
Build high-quality training datasets. Data collection strategies, cleaning pipelines, deduplication, quality filtering, format conversion (Alpaca, ShareGPT, Chat), and synthetic data generation.
Set up multi-GPU training environments. DeepSpeed, FSDP, gradient checkpointing, mixed precision, and distributed training across multiple nodes for large model fine-tuning.
Evaluate fine-tuned models rigorously. Perplexity, BLEU, ROUGE, human evaluation frameworks, LLM-as-judge, and building custom benchmarks for your domain.
Align models with human preferences. Reward modelling, PPO for RLHF, Direct Preference Optimisation (DPO), and constitutional AI approaches for safer model outputs.
Serve your models in production with vLLM and Text Generation Inference (TGI). Quantisation for deployment (GPTQ, AWQ, GGUF), batching, streaming, and cost-efficient inference at scale.
ML engineers and researchers who want to customise LLMs for specific domains and tasks
AI engineers building products that require specialised language models beyond off-the-shelf APIs
Technical leads evaluating whether to fine-tune vs. use API-based models for their organisation
Deep learning, Python, and Transformer knowledge required. You should be comfortable with PyTorch, understand Transformer architecture, and have experience training neural networks. Our Deep Learning Masterclass provides ideal preparation.
Interactive sessions with real-time Q&A and screen sharing
All sessions recorded and available for 12 months after the course
Real-world projects that build your portfolio as you learn
Personal mentoring sessions to address your specific questions
Industry-recognised certificate upon successful completion
Our instructors are seasoned practitioners with years of experience building production AI systems. They hold certifications across major cloud platforms and have trained thousands of professionals worldwide.
We provide GPU access through cloud platforms for all labs (typically A100 or equivalent). For personal experimentation, QLoRA allows fine-tuning on consumer GPUs (16GB VRAM). Budget approximately 100-200 pounds for cloud compute over 6 weeks.
We work primarily with Llama, Mistral, and Phi models of various sizes. You will learn to evaluate and select base models, so the skills transfer to any new model that is released.
Absolutely. Fine-tuning provides consistent behaviour, lower latency, reduced costs (smaller models), domain specialisation, and the ability to run models on your own infrastructure. It is complementary to RAG and prompting.
Yes, provided you use commercially-licensed base models (Llama 3, Mistral, etc.). We cover licensing considerations and help you choose models appropriate for your commercial use case.