Why Fine-Tuning Beats a Generic Base Model

▶ Watch (02:41)

Dragon Copilot supports hundreds of millions of patient interactions per year. It got there through iteration: the team started with GPT-4, found the results fell short, collected 750,000 annotated patient interactions, and kept retraining until quality met the bar. That loop is not new. Traditional MLOps ran the same cycle with classical models. Fine-tuning is how 48 percent of enterprise product leaders using generative AI today differentiate their products from competitors who all hit the same base foundation model.

Azure ML as the Fine-Tuning and Deployment Platform

▶ Watch (08:05)

Azure Machine Learning sits inside AI Foundry as the code-first training and deployment layer. It connects to structured databases or data lakes, versions datasets automatically, and supports any language stack including R alongside Python. Once a model trains, blue-green deployment controls traffic splits across model versions so teams can measure production performance directly, not just in test environments. Built-in governance tools flag fairness and bias issues before they reach users. Azure ML also trains Microsoft’s own Phi family of open-source models.

Demo: Fine-Tuning a Medical Image Model and Wiring It Into an Agent

▶ Watch (14:13)

Medical Image Insight was trained on 3.7 million scans across 14 categories but skipped gastrointestinal data entirely. The Medical Image Insight fine-tuning pipeline and agent demo (14:13) shows how to fill that gap: a versioned Azure ML pipeline distributes GI scan data across GPU clusters, logs accuracy metrics automatically, and captures full data-to-model lineage for compliance audits. Deploying the fine-tuned model generates a REST endpoint that a GPT-4o agent in AI Foundry calls directly. Upload a scan, paste in patient symptoms, and the agent returns a differential, with session context preserved between turns.

When to Use Reasoning Models and How GRPO Trains Them

▶ Watch (22:43)

Standard LLMs predict the next token. Reasoning models add a reinforcement training step that forces the model to work through a problem before answering. More reasoning tokens means more latency and compute cost. Medical, legal, and pharma applications trade that cost for accuracy and interpretability.

“except they know how to think before they respond,” — Vijay Aski

GRPO builds a reasoning model by generating multiple completions per prompt, scoring each on accuracy and format with separate reward functions, computing per-sample advantage scores, and updating model weights while KL divergence keeps the result close to the base model.

Demo: Running a Distributed GRPO Job on Azure ML

▶ Watch (33:49)

The notebook configures Qwen2.5-7B-Instruct from Hugging Face, a two-node H100 cluster, the TRL GRPO Trainer, vLLM for colocated inference, and DeepSpeed Stage 3, which shards optimizer states across GPUs and offloads optimizer computations to CPU. The GRPO reasoning model training job and profiler demo (33:49) walks through a completed job’s reward metrics and the Azure ML profiler, which aggregates all distributed logs so a single query surfaces an OOM exception. The side-by-side base vs. reasoning model comparison (45:41) shows the GRPO-trained model returning a more detailed clinical recommendation on the same GI scan.

Notable Quotes

AI isn’t just hype here. Chris Lauren · ▶ 1:21

training a reasoning model is expensive Vijay Aski · ▶ 31:04

except they know how to think before they respond, Vijay Aski · ▶ 24:00

Key Takeaways

  • Dragon Copilot iterated through 750,000 annotated patient interactions to reach production-quality fine-tuned models.
  • GRPO trains reasoning models by scoring completions on accuracy and format, then using advantage to update model weights.
  • Azure ML’s profiler surfaces OOM and distributed training errors across all GPUs and nodes from a single query.