sentence-transformers

Generates fixed-dimensional dense embeddings from variable-length text using a modular nn.Sequential pipeline (Transformer → Pooling → Dense → Normalize). The SentenceTransformer class orchestrates transformer token outputs through configurable pooling strategies (mean, max, CLS token) and optional dense projection layers, producing normalized vectors optimized for semantic similarity search. Supports asymmetric query/document encoding via Router modules for specialized model variants.

Scores or ranks text pairs by jointly encoding both sentences through a single transformer, outputting similarity scores or classification labels. The CrossEncoder class wraps AutoModelForSequenceClassification, processing concatenated sentence pairs end-to-end rather than independently encoding them, achieving higher accuracy than bi-encoder similarity comparisons at the cost of O(n) inference time per document. Includes specialized rank() method for sorting document collections by relevance to a query.

Trains models on multiple datasets simultaneously using configurable batch sampling strategies (round-robin, weighted sampling, sequential) to balance dataset contributions and prevent one dataset from dominating training. The Trainer system manages dataset loading, sampling, and loss aggregation across datasets, enabling multi-task learning and domain adaptation. Batch sampling strategies control how examples are selected from each dataset per training step, enabling flexible curriculum learning and data balancing.

Automatically generates model cards with training details, evaluation metrics, and usage instructions, and uploads trained models to Hugging Face Hub with version control and documentation. The model card system captures model architecture, training configuration, loss functions, and evaluation results, enabling reproducibility and community discovery. Hub integration enables seamless sharing, versioning, and collaborative model development with automatic README generation.

Supports prompt engineering and instruction-tuning for embedding models by allowing custom prompts to be prepended to queries and documents during encoding. The library enables task-specific prompt templates (e.g., 'Represent this document for retrieval:') that guide the model to produce task-optimized embeddings. Instruction tuning improves performance on specific tasks by conditioning embeddings on task descriptions, enabling zero-shot transfer to new tasks.

Generates sparse embeddings (high-dimensional, mostly-zero vectors) by learning per-token importance weights through a SparseEncoder architecture, enabling efficient lexical-semantic hybrid search. Unlike dense embeddings, sparse vectors preserve interpretability (which tokens matter) and integrate seamlessly with traditional BM25 retrieval systems. The architecture learns to weight tokens based on semantic relevance rather than raw term frequency, improving recall on out-of-vocabulary terms.

Fine-tunes pre-trained sentence transformers using a Trainer system supporting 40+ specialized loss functions (ContrastiveLoss, TripletLoss, MultipleNegativesRankingLoss, CosineSimilarityLoss, etc.) tailored to different training objectives. The training pipeline handles dataset preparation, batch sampling strategies, and multi-dataset training, with automatic model card generation and Hub integration for sharing trained models. Loss functions are modular and composable, enabling custom training objectives for domain-specific tasks.

Evaluates embedding and reranking models using task-specific evaluators (InformationRetrievalEvaluator, TripletEvaluator, BinaryAccuracyEvaluator, etc.) that compute standard IR metrics (NDCG, MAP, MRR, Recall@k) and classification metrics. Evaluators integrate with the Trainer system for automatic validation during training, supporting both dense and sparse model evaluation. Metrics are computed on held-out test sets and logged for model selection and hyperparameter tuning.

Loads pre-trained sentence transformer models directly from Hugging Face Hub (15,000+ models) with a single line of code, automatically downloading weights, tokenizers, and configuration. The library caches models locally and handles version management, supporting both dense (SentenceTransformer), cross-encoder (CrossEncoder), and sparse (SparseEncoder) architectures. Integration with Hub enables seamless model sharing, versioning, and community contributions.

Processes multiple texts in batches through GPU-accelerated transformers, automatically managing batch size, device placement, and memory optimization. The encode() method supports configurable batch sizes, optional tensor conversion, and multi-GPU inference via DataParallel. Batching reduces per-sample latency by 5-10x compared to single-sample inference, with automatic memory management to prevent OOM errors on large batches.

Computes pairwise semantic similarity between embeddings using multiple distance metrics (cosine similarity, Euclidean distance, dot product, Manhattan distance). The similarity() method efficiently computes similarity matrices for large embedding sets using vectorized operations, with optional normalization and threshold filtering. Supports both dense and sparse embeddings, enabling flexible similarity-based ranking and clustering.

Exports trained sentence transformer models to ONNX and OpenVINO formats for deployment on CPU-only or edge devices without PyTorch dependency. The export process converts transformer weights and pooling layers to ONNX intermediate representation, enabling inference optimization via quantization and pruning. OpenVINO export enables Intel hardware acceleration and reduced model size for embedded deployment.

Implements asymmetric encoding where queries and documents are processed through different model paths using Router modules, enabling specialized optimization for query vs. document encoding. The Router selects between different transformer configurations or pooling strategies based on input type, allowing queries to use lightweight encoders while documents use heavier models. This architecture improves retrieval quality by optimizing for the asymmetric nature of search tasks (one query vs. many documents).