Qwen News Analyzer - Fine-tuned

Model Details

• Base Model: Qwen/Qwen2.5-1.5B-Instruct
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• LoRA Rank: 64
• Training Framework: LLaMA-Factory

Tasks

1. News Details Extraction: Extracts structured information (title, keywords, summary, category, entities)
2. News Translation: Translates Arabic news to English/French with title

Training Data

• Total Samples: ~2,950
• Training: 2,700 samples
• Validation: 250 samples
• Data Generation: GPT-4o-mini labeled synthetic dataset
• Source: Arabic news articles

Training Configuration

• Epochs: 3
• Batch Size: 1 per device
• Gradient Accumulation: 4 steps
• Learning Rate: 1e-4
• Scheduler: Cosine
• Precision: bf16
• Max Sequence Length: 3,500 tokens

Usage

Performance

• Throughput (vLLM): ~630 tokens/second
• Inference Engine: vLLM with LoRA support

Limitations

• Trained specifically on Arabic news domain
• May not generalize well to other text types
• Chinese character filtering applied during generation

License