Followup improvements to GRPO job example

sdk/python/jobs/grpo/README.md

@@ -1,4 +1,40 @@
-### This directory hosts an example of running GRPO on AzureML
+### This directory hosts an example to train an Instruct model into a Reasoning model on AML using Group Relative Policy Optimization (GRPO) - a RFT technique
+
+#### Introduction:
+In this repo, one will find a notebook (launch_grpo_command_job-med-mcqa-commented.ipynb) that details an end-to-end process of fine-tuning the **Qwen2.5-7B-Instruct** model into a **reasoning model** using medical data on **Azure ML**. Qwen2.5-7B-Instruct is an instruction-tuned large language model developed by Alibaba Cloud, based on their Qwen2.5-7B foundation model. \
+It is optimized for following human instructions across a wide range of tasks, such as question answering, code generation, and language understanding. \
+In this walkthrough, one will learn how to enhance the model's reasoning capabilities using **Reinforced Fine-Tuning (RFT)** techniques, with a focus on **GRPO (**G**roup **R**elative **P**olicy **O**ptimization)**.
+
+<img src="images/agenda.png" alt="image.png" width="1000"/>
+
+#### About the GRPO Trainer
+<div style="display: flex; align-items: flex-start; gap: 32px;">
+  <div style="flex: 1;">
+    <p>The reasoning model training process typically includes three key components:</p>
+    <ul>
+      <li><strong>Sampler</strong> – Generates multiple candidate responses from the model</li>
+      <li><strong>Reward Function</strong> – Evaluates and scores each response based on criteria like accuracy or structure</li>
+      <li><strong>Trainer</strong> – Updates the model to reinforce high-quality outputs</li>
+    </ul>
+    <p>
+      In this example, the <strong>GRPO Trainer</strong> (an implementation from TRL library) is used for training Qwen2.5-7B-Instruct model into a reasoning model.
+    </p>
+    <br>
+    <p>
+      <strong>GRPO</strong> (<strong>G</strong>roup <strong>R</strong>elative <strong>P</strong>olicy <strong>O</strong>ptimization) is a reinforcement learning technique that:
+    </p>
+    <ul>
+      <li><em>Compares</em> multiple answers within a group</li>
+      <li><em>Rewards</em> the best-performing outputs</li>
+      <li><em>Penalizes</em> poor ones</li>
+      <li>Applies careful updates to <em>avoid sudden changes</em></li>
+    </ul>
+  </div>
+  <div style="flex: 1; display: flex; justify-content: center;">
+    <img src="images/training_loop.png" alt="Training Loop" style="max-width:100%; width: 600px;"/>
+  </div>
+</div> 
+
 
 #### Repo structure:
 - aml_setup.py: AzureML specific code relating to creation of dataset, model and environment. The workspace config is located here, make changes to this file before running the notebook.
@@ -10,4 +46,4 @@
     - grpo_trainer_callbacks.py: Code which converts the output models to MLflow model format. This conversation simplifies deployment as AzureML automatically generated inferencing environment and server for models in this format.
     - grpo_trainer_rewards.py: Rewards for the training, these are python functions which take an LLM response and grade it for accuracy and format.
 
-Additionally, [this video](https://youtu.be/YOm_IQt3YWw?si=5nZzyy-PZyP9XFSU&t=1344) offers an overview based on the contents of this repository.
+Additionally, [this video](https://youtu.be/YOm_IQt3YWw?si=5nZzyy-PZyP9XFSU&t=1344) offers an overview based on the contents of this repository.