135 lines
4.4 KiB
Markdown
135 lines
4.4 KiB
Markdown
---
|
|
license: apache-2.0
|
|
language:
|
|
- ar
|
|
- en
|
|
base_model: Qwen/Qwen2.5-0.5B-Instruct
|
|
tags:
|
|
- qwen2
|
|
- arabic
|
|
- iraqi
|
|
- government
|
|
- classification
|
|
- ner
|
|
- fine-tuned
|
|
- full-finetune
|
|
- goldennet
|
|
datasets:
|
|
- custom
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
model-index:
|
|
- name: GoldenNet-Qwen2.5-0.5B-Full-v1
|
|
results:
|
|
- task:
|
|
type: text-classification
|
|
name: Document Classification
|
|
metrics:
|
|
- type: loss
|
|
value: 0.3636
|
|
name: Eval Loss
|
|
---
|
|
|
|
# GoldenNet-Qwen2.5-0.5B-Full-v1
|
|
|
|
<div align="center">
|
|
<img src="https://img.shields.io/badge/Golden%20Net%20AI-Iraqi%20Gov%20NLP-gold?style=for-the-badge" alt="Golden Net AI"/>
|
|
<img src="https://img.shields.io/badge/Method-Full%20Fine--tune-red?style=for-the-badge" alt="Full"/>
|
|
<img src="https://img.shields.io/badge/Language-Arabic-green?style=for-the-badge" alt="Arabic"/>
|
|
</div>
|
|
|
|
## Model Description
|
|
|
|
**GoldenNet-Qwen2.5-0.5B-Full-v1** is a fully fine-tuned version of [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) specialized for **Iraqi Government Correspondence Processing**.
|
|
|
|
This is the **full fine-tuning** variant where all 494M parameters were trained, potentially offering the best task-specific performance.
|
|
|
|
### Tasks
|
|
1. **Document Classification** - 8 categories (طلب، شكوى، تقرير، إعلام، استفسار، دعوة، تعميم، إحالة)
|
|
2. **Named Entity Recognition** - Extracts persons, organizations, locations, dates, monetary values, laws
|
|
|
|
## Model Comparison
|
|
|
|
| Model | Method | Train Loss | Eval Loss | Training Time | Size |
|
|
|-------|--------|------------|-----------|---------------|------|
|
|
| [QLoRA-v1](https://huggingface.co/Alamori/GoldenNet-Qwen2.5-0.5B-QLoRA-v1) | 4-bit QLoRA | 0.448 | **0.2998** | 49s | 943MB |
|
|
| [LoRA-v1](https://huggingface.co/Alamori/GoldenNet-Qwen2.5-0.5B-LoRA-v1) | Standard LoRA | 0.496 | 0.3665 | 70s | 943MB |
|
|
| **Full-v1** | Full Fine-tune | **0.461** | 0.3636 | 121s | 1.9GB |
|
|
|
|
## Training Details
|
|
|
|
| Parameter | Value |
|
|
|-----------|-------|
|
|
| Base Model | Qwen/Qwen2.5-0.5B-Instruct |
|
|
| Fine-tuning Method | Full (all parameters) |
|
|
| Learning Rate | 5e-5 |
|
|
| Optimizer | AdamW 8-bit |
|
|
| Epochs | 3 |
|
|
| Batch Size | 1 (effective: 16) |
|
|
| Max Sequence Length | 1024 |
|
|
| Precision | BF16 |
|
|
| Trainable Parameters | 494M (100%) |
|
|
| Hardware | NVIDIA RTX 5070 (8GB VRAM) |
|
|
|
|
### Loss Progression
|
|
- Epoch 1: 0.983
|
|
- Epoch 2: 0.328
|
|
- Epoch 3: 0.171 (lowest among all variants!)
|
|
|
|
## Usage
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(
|
|
"Alamori/GoldenNet-Qwen2.5-0.5B-Full-v1",
|
|
device_map="auto",
|
|
torch_dtype="auto"
|
|
)
|
|
tokenizer = AutoTokenizer.from_pretrained("Alamori/GoldenNet-Qwen2.5-0.5B-Full-v1")
|
|
|
|
# Classification example
|
|
correspondence = """جمهورية العراق
|
|
وزارة التربية
|
|
العدد: 1234/ت/2025
|
|
|
|
إلى/ السيد مدير عام التعليم المحترم
|
|
|
|
م/ طلب تعيين معلمين
|
|
|
|
نرجو الموافقة على تعيين 50 معلماً.
|
|
|
|
مع التقدير"""
|
|
|
|
instruction = "صنّف المراسلة الحكومية التالية إلى إحدى الفئات: طلب، شكوى، تقرير، إعلام، استفسار، دعوة، تعميم، إحالة. أجب بصيغة JSON."
|
|
|
|
messages = [{"role": "user", "content": f"{instruction}\n\n{correspondence}"}]
|
|
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
inputs = tokenizer(text, return_tensors="pt").to(model.device)
|
|
|
|
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
|
|
print(tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True))
|
|
```
|
|
|
|
## When to Use This Model
|
|
|
|
- **Use Full-v1** when you need maximum task-specific performance and have sufficient storage/memory
|
|
- **Use QLoRA-v1** for best balance of quality and efficiency (recommended for most cases)
|
|
- **Use LoRA-v1** for comparison or when you need standard LoRA compatibility
|
|
|
|
## Related Models
|
|
|
|
- [GoldenNet-Qwen2.5-0.5B-QLoRA-v1](https://huggingface.co/Alamori/GoldenNet-Qwen2.5-0.5B-QLoRA-v1) - 4-bit quantized (best eval loss)
|
|
- [GoldenNet-Qwen2.5-0.5B-LoRA-v1](https://huggingface.co/Alamori/GoldenNet-Qwen2.5-0.5B-LoRA-v1) - Standard LoRA
|
|
|
|
## License
|
|
|
|
Apache 2.0
|
|
|
|
---
|
|
|
|
<div align="center">
|
|
<b>Developed by Golden Net AI</b><br>
|
|
<i>Empowering Iraqi Government Digital Transformation</i>
|
|
</div>
|