DeBERTa-v3-xsmall-mnli-fever-anli-ling-binary

Classifies arbitrary text into user-defined categories without task-specific fine-tuning by reformulating classification as natural language inference (NLI). The model takes input text and candidate labels, converts them into entailment hypotheses (e.g., 'This text is about [label]'), and uses the DeBERTa-v3 transformer backbone trained on MNLI, FEVER, ANLI, and LingNLI datasets to compute entailment probabilities. This approach enables dynamic label sets at inference time without retraining.

Performs entailment classification (entailment/neutral/contradiction) on English text pairs using a transformer model pre-trained on diverse NLI corpora. The model encodes premise and hypothesis as a single sequence with [CLS] token, passes through 12 DeBERTa-v3 transformer layers with disentangled attention, and outputs 3-way classification logits. Training on MNLI (formal written English), FEVER (Wikipedia claims), ANLI (adversarial examples), and LingNLI (linguistic phenomena) provides robustness across text styles and reasoning patterns.

Model is exported in multiple formats (PyTorch, ONNX, SafeTensors) enabling deployment across heterogeneous inference environments. ONNX export allows hardware-accelerated inference on CPUs, GPUs, and specialized accelerators (TPUs, NPUs) via ONNX Runtime, while SafeTensors format provides faster model loading (memory-mapped binary format) and improved security (no arbitrary code execution during deserialization). The xsmall variant (22M parameters) fits within memory constraints of edge devices and serverless functions.

Processes multiple text samples in a single inference pass by batching tokenized inputs and computing classification scores across the batch dimension. The model applies softmax normalization to logits, enabling threshold-based filtering where predictions below a confidence threshold are marked as uncertain or rejected. This capability is essential for production pipelines where confidence-based routing (e.g., escalate low-confidence samples to human review) is required.

Although trained exclusively on English NLI datasets, the model can perform limited zero-shot classification on non-English text by leveraging the multilingual tokenizer and shared transformer weights. When non-English text is tokenized and passed through the English-trained model, it relies on cross-lingual word embeddings and attention patterns learned during pre-training to generalize. Performance on non-English languages is degraded compared to English but enables zero-shot classification without language-specific fine-tuning.