Azure/azure-sdk-for-python azure-ai-finetuningsessions_1.0.0b2

on GitHub

1.0.0b2 (2026-11-12)

Features Added

• Add opt-in text-only sampling with public conversation messages and function tools, plus reference-based training using opaque sample references and scalar advantages.
• Support optional model-specific text-only reasoning_effort values low, medium, and high. Omission or None preserves the model default.
• Add async FineTuningSessionClient.attach_session(session_id) to start a heartbeat for an existing session without creating or loading it.
• Add synchronous and asynchronous local checkpoint downloads. These prepare the export, validate the complete inventory, and stream files into an empty or absent destination with exact ETag and byte-count checks. Failures retain partial output without automatic resume, deletion, or overwrite; model eligibility and authorization remain enforced by the service.

Breaking Changes

• Remove FoundryFeaturesOptInKeys.AGENT_INSIGHTS_V1_PREVIEW because Agent Insights no longer requires the obsolete preview opt-in. Remove explicit references to that constant. The fine-tuning preview header is unchanged.

Bugs Fixed

• Provide actionable checkpoint-resume guidance for terminal training-engine interruptions while preserving error codes and diagnostic identifiers.
• Preserve out-of-memory and timeout guidance for nonterminal engine errors.
• Retry sampling responses missing request_id within a bounded fault budget, then raise MalformedResponseError.
• Improve polling budgets, sampling limits, heartbeat shutdown, and scalar serialization while preserving protected-training replay safeguards.
• Use canonical /fine_tuning_sessions routes in generated operations while retaining existing Python convenience APIs.
• Preserve POSIX ownership across checkpoint directory creation and opening using protected staging and atomic no-replace publication. Reject unsupported staging explicitly and retain partial output without adopting or overwriting concurrent directories.

Other Changes

• Clarify session lifecycle, service-selected training tiers, checkpoint preparation limits, and differences between raw operations and convenience methods.
• Improve generated request-model documentation without changing supported message roles, function-only tools, or request payloads.
• Restore the standard README sections and include public test sources in source distributions while keeping archived review material excluded.
• Synchronize generation metadata with the updated public TypeSpec source without changing the Python API or runtime.