AI Foundry Lacking Support for Python Graders

[Decadz]

Technical Feedback

Hi, I have some RFT jobs where I would like to use Python graders. It would be great if these feature could be supported, as it would enable more complex training tasks.

Desired Outcome

For python graders, I would like to be able to use algorithms (from numpy, scipy, etc.) as part of the reward calculation.

Current Workaround

Not able to run python grader experiments, so instead exploring the model grader; however, this required changing the reward function design.

[amynic]

Hi @Decadz in July the Azure AI Foundry Fine Tuning team shared this blog - https://techcommunity.microsoft.com/blog/azure-ai-foundry-blog/what%E2%80%99s-new-in-azure-ai-foundry-finetuning-july-2025/4438850 - there is a section on Python Graders would this help achieve what you are trying to build?

[Decadz]

Thanks for the response @amynic,

Yes this link is helpful. However, there is a still some ambiguity in the Azure AI Foundry documentation. For example, I've been trying the following Python grader function and it fails to execute the training run:

def grade(sample, item) -> float:
    output_text = sample['output_text']
    return 1

It would be nice to have an error trace so that users can debug issues (locally all my grader functions work fine). With regard to Python graders, they tend to be more complex then other grader types, and the black box nature of AI Foundry, makes it difficult to debug.

[amynic]

This is fantastic feedback @Decadz - let me make sure we get this logged - cc @mbaiza27, @nicholasdbrady