bge-m3-zeroshot-v2.0

Classifies text into arbitrary user-defined categories without task-specific fine-tuning by leveraging XLM-RoBERTa's 111-language cross-lingual transfer capabilities. The model uses contrastive learning (trained on 53M text pairs via BGE-M3 architecture) to map input text and candidate labels into a shared embedding space, computing similarity scores to determine the most probable class. This approach enables classification across 111 languages simultaneously without retraining, using only the candidate label descriptions as guidance.

Computes dense vector embeddings for text in any of 111 languages using the BGE-M3 contrastive learning framework, enabling semantic similarity comparisons across language boundaries. The model encodes text into a 768-dimensional embedding space where semantically similar phrases cluster together regardless of language, using cosine similarity for ranking. This enables retrieval, deduplication, and clustering tasks without language-specific preprocessing or separate embedding models per language.

Supports inference via ONNX Runtime in addition to native PyTorch, enabling hardware-accelerated execution on CPUs, GPUs, and specialized inference accelerators (TPUs, NPUs). The model is distributed in both safetensors and ONNX formats, allowing deployment in resource-constrained environments (edge devices, serverless functions) with 2-5x faster inference than PyTorch on CPU-only hardware. ONNX Runtime applies graph optimization, operator fusion, and quantization-aware inference automatically.

Integrates seamlessly with the HuggingFace transformers library's zero-shot-classification pipeline, allowing single-line inference via the standard `pipeline('zero-shot-classification', model='MoritzLaurer/bge-m3-zeroshot-v2.0')` interface. The model follows transformers conventions for tokenization, model loading, and inference, enabling drop-in compatibility with existing transformers-based workflows, Hugging Face Hub model cards, and community tools without custom wrapper code.

Enables multi-label classification by computing similarity scores for all candidate labels and allowing threshold-based filtering to assign multiple labels to a single input. The model outputs a continuous similarity score (0-1) for each candidate label, enabling users to define custom confidence thresholds (e.g., assign all labels with score >0.5) rather than forcing single-label predictions. This approach supports hierarchical or overlapping classification scenarios without architectural changes.

Applies zero-shot classification to detect policy violations, harmful content, or inappropriate material across 111 languages by defining violation categories as candidate labels (e.g., 'hate speech', 'spam', 'violence') and scoring input text against them. The cross-lingual embedding space ensures consistent violation detection regardless of language, enabling moderation systems that don't require language-specific rule sets or separate classifiers per language. Similarity scores indicate violation confidence, enabling tiered moderation workflows (auto-remove >0.9, queue for review 0.5-0.9, allow <0.5).