Lack of Practical Documentation & Examples for New Users in ROLL RL Framework

Dear ROLL Team,

First of all, thank you for open-sourcing such a promising and powerful reinforcement learning framework. Your work is impressive, and the effort to integrate multiple advanced features is clearly commendable.

However, as a developer trying to get started with ROLL after reading through the README and official documentation, I found it extremely difficult to find practical, beginner-friendly information. While the documentation does a great job of highlighting the product’s features and advantages, it unfortunately lacks hands-on guidance for real-world usage.

Here are a few key questions I encountered that I believe are common to many new users:

1. **Custom Multimodal VQA Datasets**

* How can we use our own image-text (VQA-style) datasets with ROLL?
* What is the required data format or schema?

2. **Custom Reward Function**

* How can we define and register a custom reward function?
* How do we pass fields from the dataset into the reward function?
* What is the expected structure and shape of the input arguments?

3. **Using Other RL Algorithms (e.g., GSPO)**

* How can we plug in a different RL algorithm (e.g., GSPO)?
* What are the recommended steps or files to modify?

4. **Custom Logging Tools (e.g., WandB)**

* How can we switch from the default logger to use tools like Weights & Biases?

In short, while ROLL is a great project with high potential, it currently lacks practical tutorials or end-to-end examples, which makes it challenging for new users to adopt and extend.

I sincerely hope the team can prioritize improving the onboarding experience alongside new feature development. Open-source projects thrive by empowering the community, and clear, actionable documentation plays a vital role in that.

Thank you again for your excellent work—I look forward to your guidance and future updates!

Best regards,