08“LLM From Scratch” Workshop — Training a GPT-2-Scale Model Live on Azure
Colleagues ran a real training loop end to end — and launched their own PoCs right after
Role: Instructor / Author
Executive summary
Designed and delivered a hands-on workshop on building and fine-tuning LLMs/SLMs on Azure (Databricks + Fabric)—enabling colleagues and partners to start their own domain-specific LLM proofs-of-concept.
- Azure Databricks
- Microsoft Fabric
- Hugging Face Transformers
- PyTorch Lightning
- OneLake
- Microsoft Foundry
Colleagues and technical communities wanted to learn how to build LLMs/SLMs from scratch on Azure. Given my expertise, I was asked to host a workshop showing how Azure data platforms (Fabric and Databricks) can create or fine-tune LLMs, as part of an AI enablement initiative.
- Cover LLM architecture and the training/fine-tuning pipeline.
- Use Azure Databricks (distributed clusters) for preprocessing and training with Hugging Face Transformers / PyTorch Lightning.
- Use Microsoft Fabric for data management (OneLake) and experiments.
- Demonstrate training a small custom model or fine-tuning a known model on a dataset.
- Emphasize integration across Azure tools end to end.
I built slides and a live demo: data prep in Databricks on a sample text dataset, cluster setup with the right libraries, a simplified training loop for a GPT-2-scale model, and transferring the final artifact to a Fabric Lakehouse for serving. I shared a Git repo with code and instructions, delivered the session both in person and via live online meeting, and ran Q&A on performance tuning and managing models with Foundry.
The session converted LLM theory into a working artifact — attendees followed the full pipeline—preprocessing, distributed cluster setup, a GPT-2-scale training loop, transfer of the final artifact to a Fabric Lakehouse—and left with the code repo to reproduce it. Several participants started their own fine-tuning PoCs right after the workshop, which is the strongest signal a workshop can send. The materials became reusable documentation adopted by teams in other regions, positioning the group as an AI enablement reference.
Teaching reinforced my own knowledge and sharpened my ability to distill key takeaways. The live training loop was the decision that made the session work—cutting it for slideware would have kept the same content and lost the impact—so I now default to demo-first for enablement. Balancing depth and approachability kept the session valuable for varied audiences. It also revealed strong demand for AI literacy among engineers, which I continue to address through coaching.