Training on Sensitive Medical Data – Content Filtering

[Decadz]

Technical Feedback

Hi,

I am currently getting the following warning for my RFT training runs "This training file was blocked due to policy violations. Please review your data and ensure that prompts do not contain requests that violate Azure OpenAI's usage policies". For context, I am working in the medical domain training on a large text dataset (>100k instances), and the error message only comes after ~5 hours of processing.

Questions:

1. Is there a way to identify which instances are violating the content filters? I had a check through the AI Foundry UI and can't see any ways to inspecting the error trace further link.
2. Can the content filtering settings be modified (on a large corporate account)?

Desired Outcome

I would like to be able to run RFT training experiments using sensitive medical training data, or have a detailed error trace which identifies which instances are violating the content filter and why.

Current Workaround

No current work around.

[Cobb787]

Hi,
You’ve raised an important point about running RFT training with sensitive medical data. The error usually occurs when certain samples in the dataset trigger Azure’s content filters, but unfortunately, the platform doesn’t always specify which ones. A good approach is to pre-screen and clean your data, removing potentially flagged terms, and request Microsoft support for detailed logs to identify problematic entries.

For additional insights on handling sensitive data and maintaining compliance while conducting research, visit here. We provide expert knowledge about chemical substances, health-related topics, and offer laboratory reagents for both institutions and individuals. These resources can help researchers and enthusiasts maintain professional standards and improve data quality for projects like yours.