Fine-tuned LoRA adapter for Phi-3.5 Mini Instruct specialized in steel plant maintenance diagnostics, safety procedures, and equipment fault analysis.

📊 Model Details

Property	Value
Base Model	microsoft/Phi-3.5-mini-instruct
Parameters	3.8B (base) + 24MB (LoRA adapter)
Fine-tuning Method	LoRA (Low-Rank Adaptation)
LoRA Rank	8
LoRA Alpha	16
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Data	2,027 industrial maintenance Q&A pairs
Training Hardware	Apple M3 Max (MLX framework)
License	MIT

🎯 Specialization

This model is fine-tuned on real-world industrial maintenance scenarios for:

Equipment Types: Rolling Mills, Blast Furnace Blowers, Compressors, Conveyor Motors
Failure Analysis: Bearing wear, thermal overload, electrical faults, pressure system failures
Maintenance Procedures: Step-by-step repair instructions with safety protocols
Technical Specifications: Torque values, part numbers, measurement standards

🚀 Quick Start

Installation

pip install transformers torch peft

Usage (Cross-Platform)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3.5-mini-instruct",
    torch_dtype="auto",
    device_map="auto"
)

# Load fine-tuned adapter
model = PeftModel.from_pretrained(
    base_model,
    "Santhoshkumarp/phi35-maintenance-wizard-lora"
)

tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-mini-instruct")

# Generate maintenance guidance
prompt = """<|system|>
You are an expert AI maintenance engineer specializing in steel plant equipment.
<|end|>
<|user|>
Rolling Mill RM1 shows excessive vibration (12.5 mm/s) and high temperature (95°C). 
What could be the root cause and recommended actions?
<|end|>
<|assistant|>
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=500, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

print(response)

Apple Silicon (MLX - Faster)

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load(
    "microsoft/Phi-3.5-mini-instruct",
    adapter_path="Santhoshkumarp/phi35-maintenance-wizard-lora"
)

prompt = """<|system|>
You are an expert AI maintenance engineer specializing in steel plant equipment.
<|end|>
<|user|>
Rolling Mill RM1 shows excessive vibration (12.5 mm/s) and high temperature (95°C). 
What could be the root cause and recommended actions?
<|end|>
<|assistant|>
"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=500, verbose=False)
print(response)

📈 Training Details

Dataset

Size: 2,027 maintenance Q&A pairs
Sources:
- Historical incident reports
- Standard Operating Procedures (SOPs)
- Equipment maintenance manuals
- Sensor data analysis scenarios

Training Configuration

yaml

Model: microsoft/Phi-3.5-mini-instruct
Method: LoRA fine-tuning
Rank: 8
Alpha: 16
Dropout: 0.05
Learning Rate: 2e-4
Batch Size: 4
Gradient Accumulation: 4
Epochs: 3
Optimizer: AdamW
Scheduler: Cosine with warmup
Warmup Steps: 100
Max Sequence Length: 2048

Hardware

Training: Apple M3 Max (MLX framework)
Training Time: ~45 minutes
Memory: ~12GB RAM

💡 Example Outputs

Query: "Blast furnace blower shows high current (65A) and unusual noise. Diagnose the issue."

Base Phi-3.5:

"High current in a blower could indicate several issues. Check the motor, bearings, and ensure proper maintenance."

Fine-tuned Model:

"Root Cause: Motor overload due to bearing degradation or misalignment.

Immediate Actions:

STOP blower operation immediately (safety-critical equipment)

Lock out / Tag out (LOTO) - ensure zero energy state

Monitor for smoke or excessive heating

Diagnostic Steps:

Check bearing condition using vibration analysis

Measure motor winding resistance (megger test)

Inspect coupling alignment

Verify voltage balance across all phases

Parts Required: Bearing assembly (Part: BF-BRG-001), Motor contactor if damaged

Repair Procedure: Follow SOP-BF-003 for bearing replacement. Torque coupling bolts to 85 Nm."

⚠️ Limitations

Domain-Specific: Optimized for steel plant equipment, may not generalize to other industries
Safety-Critical: Always verify recommendations with qualified maintenance personnel
English Only: Trained on English-language maintenance documentation
Sensor Data: Best performance with specific numerical values (vibration, temperature, current, pressure)

🔧 System Architecture

This model is used in the Industrial Agent AI system:

User Query
    ↓
Equipment Sensor Data + Historical Context
    ↓
RAG Retrieval (Qdrant Vector DB)
    ↓
Fine-tuned Phi-3.5 Mini (this model)
    ↓
Multi-Agent Analysis
    ↓
Actionable Maintenance Plan + Citations

📄 Citation

bibtex

@misc{phi35-maintenance-wizard-2024,
  title  = {Phi-3.5 Mini Industrial Maintenance Wizard LoRA Adapter},
  author = {Santhosh Kumar P},
  year   = {2024},
  url    = {https://huggingface.co/Santhoshkumarp/phi35-maintenance-wizard-lora}
}

📜 License

MIT License - Adapter weights only. Base model license: MIT

Built for the Industrial AI Hackathon • GitHub Repository