| Property | Value |
|---|---|
| Parameter Count | 70.8M |
| License | MIT |
| Paper | DeBERTa-V3 Paper |
| Training Data | 782,357 hypothesis-premise pairs |
| Accuracy (MNLI-m) | 92.5% |
What is DeBERTa-v3-xsmall-mnli-fever-anli-ling-binary?
This is a specialized natural language inference (NLI) model built on Microsoft's DeBERTa-v3-xsmall architecture, specifically optimized for binary classification tasks. It's trained to determine whether a hypothesis is entailed by a premise or not, making it particularly effective for zero-shot classification scenarios.
Implementation Details
The model leverages advanced training on four major NLI datasets (MultiNLI, Fever-NLI, LingNLI, and ANLI), utilizing mixed precision training with carefully tuned hyperparameters including a learning rate of 2e-05 and weight decay of 0.06. The implementation features efficient batch processing and warmup optimization.
- Binary classification focus (entailment/non-entailment)
- Optimized for zero-shot applications
- Mixed precision training (FP16)
- Efficient batch processing (32 samples per device)
Core Capabilities
- High accuracy on MNLI matched (92.5%) and mismatched (92.2%) sets
- Efficient processing speed (473 texts/sec on GPU)
- Robust performance across multiple NLI datasets
- Optimized for resource-efficient deployment
Frequently Asked Questions
Q: What makes this model unique?
The model's binary classification approach and specific optimization for zero-shot classification, combined with its efficient architecture (70.8M parameters), makes it particularly suitable for practical applications where distinguishing between neutral and contradiction isn't necessary.
Q: What are the recommended use cases?
This model is ideal for zero-shot classification tasks, text entailment verification, and applications requiring binary decision-making about text relationships. It's particularly useful when computational efficiency is important while maintaining high accuracy.