Microsoft’s Azure AI Fine-Tuning Sessions Python SDK 1.0.0b2 adds text-only sampling options, session attachment and local checkpoint downloads. For developers working with fine-tuning sessions, the practical gains are more control over training samples and a guarded way to bring checkpoints onto a local machine.
Microsoft AI Watch analysis
What happened
The release notes, dated 12 November, add opt-in text-only sampling with public conversation messages and function tools, plus reference-based training using opaque sample references and scalar advantages. They also allow optional model-specific reasoning-effort settings of low, medium or high; leave the setting out, or pass None, and the model default is preserved.
Developers can now attach an asynchronous client to an existing session and start its heartbeat without creating or loading the session. The SDK also adds synchronous and asynchronous local checkpoint downloads. Those downloads validate the inventory and check ETags and byte counts while streaming files to an empty or absent destination. Failures leave partial output in place rather than automatically resuming, deleting files or overwriting a destination.
Our top picks
- Choose text-only sampling
Fine-tuning requests can use public conversation messages and function tools without the broader sampling setup. - Train from referenced samples
Opaque sample references and scalar advantages provide another supported route for preparing training data. - Set reasoning effort when the model supports it
Choose low, medium or high, or omit the value to keep the model’s default. - Attach to an existing session
The asynchronous client can start a heartbeat without creating or loading the session. - Download checkpoints with inventory checks
Local downloads verify file counts and ETags, and avoid automatically overwriting an existing destination. - Get clearer interruption guidance
The SDK adds actionable checkpoint-resume guidance for terminal training-engine interruptions while preserving error codes and diagnostic identifiers.
Why it matters
This is a useful set of changes for teams building and operating fine-tuning workflows: more sample options, a way to reconnect to existing sessions, and local checkpoint handling with explicit safeguards. The download behaviour is particularly worth noting: a failed transfer can leave partial output, but the SDK will not quietly resume it or clean it up for you.
There is also a breaking change. The obsolete AGENTINSIGHTSV1PREVIEW opt-in constant and its references have been removed because Agent Insights no longer requires that preview opt-in. The notes say the fine-tuning preview header is unchanged.
Our read
This is more than a version number nudging its way into a changelog. The new controls and checkpoint safeguards give Python developers concrete workflow options, while the removed opt-in is a small but real compatibility detail. If you use the SDK, check for references to the removed constant and make sure your checkpoint-handling code accounts for partial downloads. Sensible engineering, with fewer surprises hiding in the staging directory.
What to watch
- Whether the new sampling options and reasoning-effort settings are supported by the models you use.
- How your workflow handles partial checkpoint output after a failed download.
- Whether removing the obsolete opt-in constant affects any code that still references it.
Discussion spark: Would you rather the SDK automatically resume or clean up a failed checkpoint download, or is leaving partial files untouched the safer default?
Sources and evidence
- azure-ai-finetuningsessions1.0.0b2 (13 November 2026, 01:36 UTC)
not affiliated with or endorsed by Microsoft