GoldenNet-Qwen2.5-3B-QLoRA-v1

Model Description

GoldenNet-Qwen2.5-3B-QLoRA-v1 is a QLoRA fine-tuned version of Qwen/Qwen2.5-3B-Instruct specialized for Iraqi Government Correspondence Processing.

This is the larger 3B model variant, offering improved performance over the 0.5B models while remaining efficient for edge deployment.

Tasks

Document Classification - 8 categories (طلب، شكوى، تقرير، إعلام، استفسار، دعوة، تعميم، إحالة)
Named Entity Recognition - Extracts persons, organizations, locations, dates, monetary values, laws

Model Comparison

Model	Size	Method	Train Loss	Eval Loss	Training Time
0.5B-QLoRA-v1	0.5B	QLoRA	0.448	0.2998	49s
0.5B-LoRA-v1	0.5B	LoRA	0.496	0.3665	70s
0.5B-Full-v1	0.5B	Full	0.461	0.3636	121s
3B-QLoRA-v1	3B	QLoRA	0.396	0.2521	14min

Training Details

Parameter	Value
Base Model	Qwen/Qwen2.5-3B-Instruct
Fine-tuning Method	QLoRA (4-bit quantization + LoRA)
Quantization	4-bit (bitsandbytes)
LoRA Rank	64
LoRA Alpha	128
LoRA Dropout	0.05
Learning Rate	1e-4
Epochs	3
Batch Size	1 (effective: 16)
Max Sequence Length	1024
Precision	BF16
Total Parameters	3.2B
Trainable Parameters	119.7M (3.7%)
Hardware	NVIDIA RTX 5070 (8GB VRAM)

Loss Progression

Epoch 0.5: 1.143
Epoch 1.0: 0.462
Epoch 1.5: 0.295
Epoch 2.0: 0.244
Epoch 2.5: 0.192
Epoch 3.0: 0.181 (final)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Alamori/GoldenNet-Qwen2.5-3B-QLoRA-v1",
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Alamori/GoldenNet-Qwen2.5-3B-QLoRA-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 3B-QLoRA-v1 for best accuracy when you have sufficient VRAM (~6GB for inference)
Use 0.5B-QLoRA-v1 for fast inference on constrained hardware
The 3B model shows ~16% improvement in eval loss over the 0.5B models