PersianMind-v1.0
| Property | Value |
|---|---|
| Developers | Pedram Rostami, Ali Salemi, Mohammad Javad Dousti |
| Model Size | 13.7GB (BF16) |
| Languages | Persian, English |
| License | CC BY-NC-SA 4.0 (Non-commercial use only) |
| Paper | arXiv:2401.06466 |
What is PersianMind-v1.0?
PersianMind-v1.0 is a state-of-the-art cross-lingual language model specifically designed for Persian and English language processing. Developed by researchers at the University of Tehran, it achieves remarkable performance on Persian language tasks and demonstrates capabilities comparable to GPT-3.5-turbo in Persian reading comprehension.
Implementation Details
The model offers flexible deployment options with support for different quantization levels (BF16, INT8, and NF4), making it adaptable to various computational resources. The model can be easily integrated using the Hugging Face Transformers library and supports both CPU and GPU execution.
- BF16 version: 13.7GB size, 25.35 tokens/sec
- INT8 version: 7.2GB size, 11.36 tokens/sec
- NF4 version: 3.9GB size, 24.36 tokens/sec
Core Capabilities
- Achieves SOTA results on Persian subset of Belebele benchmark
- Excels in ParsiNLU multiple-choice QA tasks
- Strong performance in Persian-English and English-Persian translation
- Comparable performance to GPT-3.5-turbo in Persian reading comprehension
Frequently Asked Questions
Q: What makes this model unique?
PersianMind-v1.0 is specifically optimized for Persian-English cross-lingual tasks, offering state-of-the-art performance while maintaining efficiency through various quantization options. It's particularly notable for achieving high accuracy on the Belebele benchmark (73.9% for BF16) and strong translation capabilities as measured by the Comet metric.
Q: What are the recommended use cases?
The model is ideal for Persian language processing tasks including reading comprehension, question-answering, and bi-directional translation between Persian and English. It's particularly suitable for academic and research applications, given its non-commercial license.
