nassimjp/Maral-7B-alpha-1

Maral-7B-alpha-1: A Persian-Focused LLM

Maral-7B-alpha-1 is a 7 billion parameter large language model built upon the Mistral architecture, specifically fine-tuned for the Persian language. Developed by Muhammadreza Haghiri and Mahi Mohrechi, this model represents a significant effort to enhance AI capabilities for Persian speakers.

Key Capabilities and Features

• Persian Language Specialization: Primarily trained on an Alpaca Persian dataset, making it proficient in generating and understanding Persian text.
• Multilingual Support: Inherits English language capabilities from its Mistral base, allowing for responses in both Persian and English.
• Instruct-Following: Designed to follow instructions using the Guanaco prompt format (### Human: <prompt>\n### Assistant: <answer>).
• Resource Optimization: Supports 4-bit and 8-bit quantization for inference on consumer hardware or smaller GPUs, with options for PEFT and bitsandbytes integration.

Known Issues and Considerations

• Hallucinations: The model can produce highly imaginative but potentially misinforming answers, particularly with reasoning tasks in Persian. This is an area for future improvement through better datasets and training (e.g., DPO).
• Repetitive Output: May generate repetitive text; a temporary solution is to keep the generation temperature between 0.5 and 0.7.
• Grammar Quality: While capable of GPT-3.5 level grammar in Persian, it may still exhibit issues that can be addressed with further training.

Use Cases

This model is particularly well-suited for applications requiring natural language understanding and generation in Persian, such as chatbots, content creation, and language translation, especially where a balance between Persian and English capabilities is beneficial.