| Property | Value |
|---|---|
| Parameter Count | 178M |
| Model Type | Natural Language Inference |
| Language | Russian |
| Downloads | 152,044 |
| Task Type | Zero-Shot Classification |
What is rubert-base-cased-nli-threeway?
This is a specialized Russian language model based on DeepPavlov/rubert-base-cased, fine-tuned specifically for natural language inference tasks. It can determine the logical relationship between two text segments, classifying them into three categories: entailment, contradiction, or neutral. The model has been trained on a comprehensive collection of translated NLI datasets and achieves strong performance across various benchmark tests.
Implementation Details
The model is built on the BERT architecture and has been optimized for Russian language processing. It utilizes advanced transformer technology and implements a three-way classification system, making it particularly effective for complex language understanding tasks.
- Built on DeepPavlov/rubert-base-cased architecture
- Supports both direct NLI tasks and zero-shot classification
- Trained on multiple translated datasets including JOCI, MNLI, MPE, SICK, and SNLI
- Achieves high ROC AUC scores across various evaluation sets
Core Capabilities
- Three-way classification (entailment, contradiction, neutral)
- Zero-shot text classification functionality
- Sentiment analysis through hypothesis testing
- High performance on Russian language tasks
- Support for both PyTorch and ONNX formats
Frequently Asked Questions
Q: What makes this model unique?
The model stands out for its specialized three-way classification capability in Russian language NLI tasks, offering better granularity compared to binary classification models. It demonstrates superior performance across multiple evaluation datasets and supports zero-shot classification tasks.
Q: What are the recommended use cases?
The model is ideal for natural language inference tasks in Russian text, sentiment analysis, text classification, and logical relationship detection between text pairs. It's particularly useful for applications requiring zero-shot classification capabilities where training data might be limited.