phi3-industrial-anomaly

🏭 Phi-3 Mini Fine-tuned for Industrial Anomaly Detection

<div align="center">

</div>

Fine-tuned version of Microsoft's Phi-3-mini-4k-instruct using QLoRA (Quantized Low-Rank Adaptation) for industrial IoT anomaly detection and interpretable diagnostics.

📋 Model Description

This model specializes in analyzing industrial sensor data and network telemetry to detect anomalies, identify potential security threats, and provide actionable insights for industrial automation systems.

Key Features:

• 🎯 Industrial anomaly classification
• 🔒 Security threat detection
• 📊 Sensor data interpretation
• 🚨 Real-time diagnostic recommendations
• 💡 Explainable AI responses

🔧 Training Details

Base Model

• Architecture: Phi-3-mini-4k-instruct (3.8B parameters)
• Context Length: 4096 tokens
• Quantization: 4-bit NF4 with double quantization

Fine-tuning Configuration

• Method: QLoRA (Quantized Low-Rank Adaptation)
• LoRA Rank: 32
• LoRA Alpha: 64
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Dropout: 0.05

Training Parameters

• Epochs: 5
• Batch Size: 4 per device
• Gradient Accumulation: 4 steps (effective batch size: 16)
• Learning Rate: 2e-5
• Optimizer: paged_adamw_8bit
• Scheduler: Cosine with warmup (100 steps)
• Mixed Precision: BF16

Dataset

• Name: Edge-Industrial-Anomaly-Phi3
• Training Samples: 10,749
• Evaluation Samples: 1,195
• Format: Conversational (user/assistant format)

📊 Evaluation Results

🚀 Usage

Using Transformers (Recommended)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "YOUR_USERNAME/phi3-industrial-anomaly",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "YOUR_USERNAME/phi3-industrial-anomaly",
    trust_remote_code=True
)

# Prepare input
prompt = """<|user|>
Sensor Readings: Temperature: 95°C, Vibration: 5.8 m/s, Pressure: 120 kPa, Flow Rate: 6.2 L/min
<|end|>
<|assistant|>"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate response
outputs = model.generate(
    **inputs,
    max_new_tokens=150,
    temperature=0.7,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Using PEFT (Load Adapters Only)

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
import torch

# Load model with LoRA adapters
model = AutoPeftModelForCausalLM.from_pretrained(
    "YOUR_USERNAME/phi3-industrial-anomaly",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "YOUR_USERNAME/phi3-industrial-anomaly",
    trust_remote_code=True
)

Example Prompts

Network Security Analysis:

<|user|>
Network Telemetry: Arp.Opcode: 0.0, Icmp.Checksum: 0.0, Suspicious packet patterns detected
<|end|>
<|assistant|>

Sensor Diagnostics:

<|user|>
Sensor Readings: Temperature: 110°C, Vibration: 7.2 m/s, Pressure: 85 kPa, Flow Rate: 3.1 L/min
<|end|>
<|assistant|>

🎯 Use Cases

• Industrial IoT Monitoring: Real-time anomaly detection in manufacturing plants
• Predictive Maintenance: Early warning systems for equipment failure
• Security Operations: Network intrusion detection in OT/IT environments
• Edge Deployment: Lightweight inference on industrial gateways and edge devices
• Smart Manufacturing: Quality control and process optimization

🛠️ Edge Deployment

Model Formats Available

• PyTorch (this repo): Full model for transformers
• GGUF: For llama.cpp and edge devices (see releases)
• ONNX: For optimized inference (convert with Optimum)

Hardware Requirements

• GPU Inference: 8GB+ VRAM (with quantization)
• CPU Inference: 16GB+ RAM
• Edge Devices: Compatible with Jetson Nano, Raspberry Pi 5, Industrial PCs

📈 Performance Considerations

• Quantization: Model uses 4-bit quantization for efficient memory usage
• Inference Speed: ~14.7 samples/second on NVIDIA RTX GPUs
• Context Window: 4096 tokens (sufficient for detailed sensor logs)
• Generation: Typical response time 2-5 seconds on GPU

⚠️ Limitations

• Model may require domain-specific fine-tuning for your specific industrial environment
• Best performance with sensor data in the format seen during training
• Evaluation accuracy (54.51%) suggests room for improvement with more training epochs
• Not suitable for safety-critical decisions without human oversight

🔄 Version History

• v1.0 (2026-01-06): Initial release
  • 5 epochs of QLoRA fine-tuning
  • LoRA rank 32, alpha 64
  • Trained on Edge-Industrial-Anomaly-Phi3 dataset

📄 Citation

If you use this model, please cite:

@misc{phi3-industrial-anomaly-2026,
  author = {Your Name},
  title = {Phi-3 Mini Fine-tuned for Industrial Anomaly Detection},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/YOUR_USERNAME/phi3-industrial-anomaly}}
}

📜 License

This model is released under the MIT License. The base Phi-3 model is subject to Microsoft's Phi-3 license.

🙏 Acknowledgments

• Microsoft Research: For the Phi-3-mini-4k-instruct base model
• Hugging Face: For the transformers and PEFT libraries
• Dataset: ssam17/Edge-Industrial-Anomaly-Phi3

📞 Contact

For questions, issues, or collaboration opportunities, please open an issue in the repository or contact the model author.

---

<div align="center"> Built with ❤️ for Industrial AI </div>