Training Hub v0.10.0

Training Hub v0.10.0

Minor release adding a new algorithm for contrastive embedding fine-tuning, E2E test coverage for the unified callback abstraction, and documentation updates.

Highlights

• Embedding SFT: New embedding_sft() entrypoint for contrastive fine-tuning of sentence embedding models (e.g. all-MiniLM-L6-v2) via the sentence-transformers backend. Supports batch-all/batch-hard triplet and MNRL losses, custom loss functions, batch sampler auto-selection, and evaluation datasets — designed for semantic routing and embedding classification workloads
• Routing demo notebook: New routing_demo.ipynb walkthrough showing semantic routing with embedding_sft — baseline (untrained) router vs. fine-tuned router over a 4-class workload with full stats
• Unified callback E2E tests: 50 new tests covering all 49 test cases from the RHAISTRAT-1256 test plan for the TrainingHubCallback abstraction across InstructLab, Mini-Trainer, and Unsloth adapters

New Features

• Embedding SFT algorithm (embedding_sft()) with sentence-transformers backend, triplet/MNRL losses, custom loss support, and batch sampler auto-selection (#139)
• New embedding dependency extra (sentence-transformers>=5.0, with 5.0–6.x import compatibility) (#139)
• Semantic routing demo notebook (examples/notebooks/routing_demo.ipynb) (#139)
• Algorithm, API, and sidebar documentation for embedding_sft (#139)

Other

• Added E2E test suite for the unified callback abstraction (RHOAIENG-79856) — 50 tests, all runnable without a GPU (#143)
• Added GEPA to the support matrix docs (#144)

What's Changed

• docs: add GEPA to the support matrix by @Maxusmusti in #144
• Add embedding_sft algorithm for contrastive embedding fine-tuning by @Maxusmusti in #139
• Add E2E tests for unified callback abstraction by @Sridhar1030 in #143

Full Changelog: v0.9.8...v0.10.0