CrossEncoder

This release introduces Multi-Vector Embedding models (ColBERT-style) as a new model type, enhancing retrieval capabilities. It also modernizes dependencies, fixes silent scoring bugs, and improves performance.

This minor release focuses on correctness and robustness, fixing silent scoring bugs related to chat-template truncation and improving stability for hard-negative mining and GIST losses. It also introduces MPS support and restores TSDAE functionality on recent `transformers` versions.

This release introduces the `train-sentence-transformers` Agent Skill and two new training losses, `EmbedDistillLoss` and `ADRMSELoss`, alongside numerous robustness and correctness improvements.

This release introduces major first-class multimodal support for SentenceTransformer and CrossEncoder, enabling unified processing of text, images, audio, and video. The CrossEncoder architecture has also been fully modularized, enabling support for generative rerankers.

Version 5.2.0 adds multiprocessing to CrossEncoder, multilingual NanoBEIR support, similarity scores in hard‑negative mining, and updates for Transformers 5 while deprecating Python 3.9 and the old `n-tuple-scores` format.

Version 4.1.0 adds ONNX and OpenVINO backends for CrossEncoder, a new `backend` argument, and utilities for model optimization, while remaining backward compatible.

Version 4.0.2 introduces safer max-length handling for CrossEncoder models and improves distributed training device placement, while fixing typing, FSDP, and documentation issues.

Version 4.0.1 introduces a complete overhaul of the CrossEncoder training pipeline with a new `CrossEncoderTrainer`, dataset‑based inputs, multi‑GPU and bf16 support, and many training‑related enhancements, while keeping inference unchanged.

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.