| Property | Value |
|---|---|
| Base Model | facebook/mbart-large-50 |
| Task | Text Summarization |
| Language | Persian (Farsi) |
| Hugging Face | eslamxm/mbart-finetuned-fa |
What is mbart-finetuned-fa?
mbart-finetuned-fa is a specialized version of Facebook's mBART-large-50 model, fine-tuned specifically for Persian language text summarization. The model demonstrates strong performance metrics, including a ROUGE-1 score of 44.07 and a BERTScore of 78.95, making it particularly effective for Persian text processing tasks.
Implementation Details
The model was trained using a carefully tuned configuration with Adam optimizer and linear learning rate scheduling. Key training parameters include a learning rate of 0.0005, batch size of 32 (accumulated), and 5 epochs of training with label smoothing factor of 0.1.
- Trained on the pn_summary dataset
- Implements gradient accumulation with 8 steps
- Uses warmup steps of 250
- Achieves a generation length average of 41.7 tokens
Core Capabilities
- ROUGE-1: 44.07 performance
- ROUGE-2: 25.81 score
- ROUGE-L: 38.96 score
- BERTScore: 78.95
- Optimized for Persian text summarization tasks
Frequently Asked Questions
Q: What makes this model unique?
The model's specialization in Persian language text summarization, combined with its strong performance metrics, makes it particularly valuable for processing Farsi content. The careful optimization of training parameters and the use of the robust mBART-large-50 architecture as its foundation contribute to its effectiveness.
Q: What are the recommended use cases?
This model is best suited for Persian text summarization tasks, particularly when working with content that requires maintaining semantic accuracy while generating concise summaries. It's ideal for applications in news summarization, document processing, and content optimization in Farsi.