SilentStorm99/opspilot-phi3-lora-v6

OpsPilot Phi-3 v6 LoRA

Fine-tuned Phi-3-mini model for IT incident classification with 99-100% accuracy.

🎯 Model Details

• Base Model: microsoft/Phi-3-mini-4k-instruct
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• LoRA Config: r=16, alpha=32, dropout=0.05
• Task: 6-category incident classification
• Accuracy: 99-100% on test set
• Training Samples: 26 real-world incidents
• Training Epochs: 20

📊 Performance

Metric	Value
Test Accuracy	99-100%
Categories	6
Average Latency	~17s (with RAG)
Model Size	LoRA adapters only (~50MB)

🏷️ Categories

1. Application - Application-level errors and crashes
2. Database - Database connectivity and performance issues
3. Infrastructure - Hardware and infrastructure problems
4. Network - Network connectivity and routing issues
5. Performance - System performance degradation
6. Security - Security incidents and breaches

🚀 Usage

With PEFT (Recommended)

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapters
model = PeftModel.from_pretrained(
    base_model,
    "SilentStorm99/opspilot-phi3-lora-v6"
)

tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")

# Classify incident
prompt = """<|system|>You are an IT incident classification assistant.<|end|>
<|user|>Classify this IT incident:
Database connection timeout errors. Pool size at maximum.
<|end|>
<|assistant|>"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

In Production (OpsPilot Server)

This model is used in the OpsPilot production server with:

• RAG (Retrieval-Augmented Generation) for knowledge enhancement
• MastraAI multi-agent escalation for complex cases
• Three-tier intelligent routing based on confidence

📚 Training Details

Dataset

• 26 carefully curated incident examples
• Natural distribution: Application (12), Infrastructure (4), Database (3), Performance (3), Security (2), Network (2)
• RAG-aware training with internal documentation

Hyperparameters

• Learning Rate: 1e-4
• Batch Size: 1
• Gradient Accumulation: 4
• Epochs: 20
• Max Length: 1024 tokens
• LoRA Rank (r): 16
• LoRA Alpha: 32
• LoRA Dropout: 0.05

Training Environment

• GPU: CUDA-enabled (RTX 4090 or similar)
• Training Time: ~2 hours
• Framework: Transformers + PEFT + BitsAndBytes

🔧 Model Architecture

Base: Phi-3-mini-4k-instruct (3.8B parameters)
  ↓
LoRA Adapters (16.8M trainable parameters)
  ↓
Sequence Classification Head
  ↓
6 Categories Output

📖 Citation

@misc{opspilot-phi3-v6,
  title={OpsPilot Phi-3 v6 LoRA for Incident Classification},
  author={SilentStorm99},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/SilentStorm99/opspilot-phi3-lora-v6}}
}

📝 License

MIT License - See repository for details

🔗 Related

• Project Repository: https://github.com/o0SilentStorm0o/OpsPilot
• Base Model: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
• PEFT Library: https://github.com/huggingface/peft

🙏 Acknowledgments

Built with Microsoft Phi-3, HuggingFace Transformers, and PEFT.