QADataGenerator doesn't generate QA pairs in non-English languages

[pamelafox]

• Package Name: azure-ai-generative
• Package Version: 1.0.0b2
• Operating System: Linux/Debian
• Python Version: 3.11

Describe the bug

When we use the QADataGenerator(model_config=openai_config) with Brazilian Portuguese texts, the QA pairs generated are always in English.

I believe that is a known issue, but it needs to be documented clearly, as customers expect LLM-based tools to work in non-English languages.

To Reproduce

Pass in Portuguese text to this code:

    result = qa_generator.generate(
        text=text,
        qa_type=QAType.LONG_ANSWER,
        num_questions=2,
    )

And get English answers instead.

Expected behavior

Portuguese answers.

[l0lawrence]

Hi @pamelafox thank you for the feedback, forwarding your request to @azureml-github and they will get back to you asap.

[diondrapeck]

@pamelafox Thank you for bringing this to our attention. We will update the reference documentation to make the language limitations clear.

[EMjetrot]

Hi everyone - Is there any progress on making QADataGenerator multilingual? I would be very much appreciated if that was possible :)

[diondrapeck]

@EMjetrot - the azure-ai-generative package was deprecated earlier this year, so no updates will be made to any of the classes. Instead, it's recommended that you use the azure-ai-evaluation package. The analogous class is in that library is the Simulator (and its subclasses) which does have support for multiple non-English languages. Here's an example.