| Property | Value |
|---|---|
| Parameters | 306M |
| License | Apache 2.0 |
| Architecture | BART Distilled |
| Training Data | CNN/DailyMail Dataset |
What is distilbart-cnn-12-6?
DistilBART CNN 12-6 is a compressed version of the BART-large-cnn model, designed for efficient text summarization. It maintains nearly identical performance to its larger counterpart while offering improved inference speed. The model name indicates its architecture: 12 encoder layers and 6 decoder layers.
Implementation Details
This model is implemented using the BartForConditionalGeneration architecture and achieves a 1.24x speedup compared to the baseline BART-large-cnn model. With 306M parameters, it achieves impressive ROUGE scores (ROUGE-2: 21.26, ROUGE-L: 30.59) that are comparable to the full model.
- Optimized inference time of 307ms
- Maintains 99% of the original model's performance
- Designed for production deployment with reduced computational requirements
Core Capabilities
- Text summarization optimized for news articles
- Efficient processing of long documents
- Compatible with both PyTorch and JAX frameworks
- Suitable for deployment on inference endpoints
Frequently Asked Questions
Q: What makes this model unique?
This model achieves an optimal balance between performance and efficiency, using knowledge distillation to compress BART while maintaining high-quality summarization capabilities.
Q: What are the recommended use cases?
The model is particularly well-suited for news article summarization, content condensation, and production environments where efficiency is crucial without compromising quality.