ticket-classifier-qwen25-7b-fp16

IT Ticket Classifier - Qwen2.5 7B FP16 Quantized

Model Details

• Base Model: unsloth/Qwen2.5-7B-Instruct
• Fine-tuning: LoRA adapter merged
• Quantization: FP16 (Half Precision)
• Date: 2026-06-05 11:00:29
• Original Accuracy: 78.72%
• Expected Accuracy (FP16): 78.6-78.7% (99% retained)

File Size

• Original (FP32): ~15 GB
• Quantized (FP16): ~7.5 GB
• Compression: 50%

Usage with Ollama

Method 1: From Hugging Face (Recommended)

# Clone the repository locally
git clone https://huggingface.co/HoangPhu/ticket-classifier-qwen25-7b-fp16
cd ticket-classifier-qwen25-7b-fp16

# Create Ollama model
ollama create ticket-classifier-fp16 -f Modelfile

# Test the model
ollama run ticket-classifier-fp16 "My VDI is not connecting"

Method 2: Using Ollama API

import requests

url = "http://localhost:11434/api/generate"
payload = {
    "model": "ticket-classifier-fp16",
    "prompt": "My laptop is slow",
    "stream": False
}

response = requests.post(url, json=payload)
print(response.json()['response'])

Service Categories

The model classifies tickets into 21 categories:

• Account Troubleshooting
• Application Troubleshooting
• Client Troubleshooting
• Create New VDI
• Deploy Application
• Enable VDI
• Endpoint Security
• FPT Chat Deployment
• ISMS Support
• Infrastructure Support
• Install/Remove Software
• Network Port Configuration
• Network Troubleshooting
• Office Installation
• Other / Manual Review
• PASSWORD_RESET
• Register Application Entra ID
• Remove VDI
• Reset MFA O365
• Upgrade VDI
• VDI Troubleshooting

Model Performance

• Accuracy: 78.72%
• High-Confidence (≥30%) Accuracy: 100%
• Recommended Production Threshold: 30%

Deployment Strategy

Recommended: Hybrid Confidence-Based Routing

If confidence >= 30%:
  → Auto-route to service team
Else:
  → Manual review required

This ensures high accuracy for auto-routed tickets while maintaining oversight for uncertain cases.

Requirements

• Ollama (v0.1.0+)
• At least 16 GB RAM for CPU inference
• Or GPU (recommended for faster inference)

Inference Speed

• CPU (Intel i7): 5-10 seconds per ticket
• GPU (NVIDIA): 0.5-2 seconds per ticket

License

Same as source model (Qwen2.5-7B-Instruct)

Quantization Details

This model was converted from FP32 (32-bit float) to FP16 (16-bit float):

• Minimal quality loss (~1%)
• 50% size reduction
• Faster loading times
• Compatible with most hardware