●Change8

v3.4.0

Breaking Changes
📦 sentence-transformersView on GitHub →
⚠ 1 breaking✨ 5 features🐛 10 fixes🔧 14 symbols

Summary

Version 3.4.0 fixes a major memory‑leak issue, adds compatibility between cached losses and MatryoshkaLoss, introduces several new features, and resolves numerous bugs.

⚠️ Breaking Changes

  • SentenceTransformerModelCardData no longer stores a reference to SentenceTransformerTrainer, breaking the previous circular reference; code that accessed model_card_data.trainer must be updated to obtain the trainer elsewhere.

Migration Steps

  1. If your code accessed SentenceTransformerModelCardData.trainer, refactor to keep a separate reference to the trainer.
  2. Reinstall or upgrade to sentence-transformers==3.4.0 using pip.
  3. Review any custom module saving logic to ensure kwargs are correctly persisted.

✨ New Features

  • Added compatibility to combine Cached losses (CachedMultipleNegativesRankingLoss, CachedGISTEmbedLoss, CachedMultipleNegativesSymmetricRankingLoss) with MatryoshkaLoss.
  • Added Matthews Correlation Coefficient metric to BinaryClassificationEvaluator.
  • Added a margin parameter to TripletEvaluator.
  • Model cards now include dataset information in expandable sections when many datasets are present.
  • Enabled multi‑GPU and CPU multi‑process support for util.mine_hard_negatives.

🐛 Bug Fixes

  • Fixed NoDuplicatesBatchSampler producing identical subsequent batches.
  • Fixed crash when using old‑style model.fit() with write_csv on an evaluator.
  • Converted evaluator output types from np.float to native float.
  • Allowed specifying revision and cache_dir when loading PEFT Adapter models.
  • Fixed CrossEncoder lazy placement on incorrect device and made it respect model.to().
  • Correctly saved custom module kwargs in modules.json (e.g., for jina‑embeddings‑v3).
  • Fixed HfArgumentParser(SentenceTransformerTrainingArguments) crash caused by prompts typing.
  • Fixed breaking change in PyLate when loading modules.
  • Raised error for empty dataset list in NanoBEIREvaluator.
  • Updated Sphinx version and switched from recommonmark to myst-parser in docs.

🔧 Affected Symbols

SentenceTransformerModelCardDataSentenceTransformerTrainerlosses.CachedMultipleNegativesRankingLosslosses.CachedGISTEmbedLosslosses.CachedMultipleNegativesSymmetricRankingLosslosses.MatryoshkaLossBinaryClassificationEvaluatorTripletEvaluatorutil.mine_hard_negativesCrossEncoderHfArgumentParserSentenceTransformerTrainingArgumentsNanoBEIREvaluatorPyLate