GoldenNet-Qwen2.5-0.5B-Full-v1 Overview

This model is a fully fine-tuned version of Qwen/Qwen2.5-0.5B-Instruct, developed by Golden Net AI. It is specifically optimized for Iraqi Government Correspondence Processing in Arabic, with all 494 million parameters trained for maximum task-specific performance.

Key Capabilities

• Document Classification: Categorizes government correspondence into 8 distinct types: طلب (request), شكوى (complaint), تقرير (report), إعلام (notification), استفسار (inquiry), دعوة (invitation), تعميم (circular), and إحالة (referral).
• Named Entity Recognition (NER): Extracts critical entities such as persons, organizations, locations, dates, monetary values, and laws from Arabic text.

Training and Performance

The Full-v1 variant underwent full fine-tuning with a learning rate of 5e-5 over 3 epochs, achieving the lowest training loss (0.171) among its related models. While its evaluation loss (0.3636) is slightly higher than the QLoRA-v1 variant, it represents the most comprehensive fine-tuning approach. The model was trained on an NVIDIA RTX 5070 with BF16 precision and a max sequence length of 1024.

When to Use This Model

• Optimal Performance: Choose Full-v1 when the highest possible task-specific accuracy is required for Iraqi government correspondence processing and computational resources (storage/memory) are not a limiting factor.
• Specialized Arabic NLP: Ideal for applications demanding precise classification and entity extraction from formal Arabic government documents.