Edge-Industrial-Anomaly-Phi3: A Curated Dataset for SLMs

This dataset is a curated collection of industrial sensor data formatted specifically for Small Language Models (SLMs) like Phi-3. It merges three high-value industrial domains into a unified "Natural Language Reasoning" format to move beyond simple binary classification.

🚀 Purpose

Standard anomaly detection uses CSVs and Scikit-Learn. This dataset enables Generative Anomaly Detection, where a model like Phi-3 can explain why an anomaly is happening based on sensor drifts and thresholds.

Key Innovation

Instead of outputting 0 or 1, the fine-tuned model generates:

• Status classification (NORMAL/ANOMALY/SECURITY ANOMALY)
• Root cause analysis in natural language
• Actionable recommendations for operators

This approach is ideal for edge deployment where human operators need interpretable AI assistance, not just binary alerts.

👥 Authors

Name	Email	GitHub	ORCID	Google Scholar	ResearchGate
Samaresh Kumar Singh	ssam3003@gmail.com	GitHub	ORCID	Google Scholar	—
Joyjit Roy	joyjit.roy.tech@gmail.com	GitHub	ORCID	Google Scholar	ResearchGate

📊 Included Data Sources

The dataset is derived and transformed from three complementary Kaggle datasets:

1. Industrial Fault Detection

• Source: Kaggle - Industrial Fault Detection Dataset
• Domain: Manufacturing equipment monitoring
• Features: Temperature, Pressure, Vibration, Flow Rate
• Use Case: Real-time equipment health monitoring

2. Edge-IIoTset Cyber Security

• Source: Kaggle - Edge-IIoTset
• Domain: IoT/IIoT network security
• Features: Network telemetry, packet statistics, protocol metrics
• Use Case: Intrusion detection for edge devices

3. Machine Failure Prediction

• Source: Kaggle - Predictive Maintenance
• Domain: Predictive maintenance for rotational machinery
• Features: Rotational speed, torque, tool wear, temperature
• Use Case: Preventing catastrophic equipment failures

🛠 Format: Phi-3 Chat Template

Each entry is formatted in the Phi-3 chat structure using special tokens:

<|user|>
[Sensor data summary]
<|end|>
<|assistant|>
[Status and natural language reasoning]
<|end|>

Example Entry

Input (User):

Sensor Readings: Temperature: 82°C, Vibration: 3.8 m/s, Pressure: 112 kPa, Flow: 16.5 L/min

Output (Assistant):

Status: ANOMALY. Reason: Critical overheating detected. Temperature (82°C) exceeds the safety threshold of 75°C. Excessive vibration detected (3.8 m/s). May indicate mechanical imbalance. Immediate inspection required.

📂 Dataset Structure

The dataset consists of three JSONL files:

├── industrial_faults_phi3.jsonl    (~10K entries)
├── edge_iiot_phi3.jsonl            (~10K entries)
└── predictive_maint_phi3.jsonl     (~10K entries)

Each JSONL file contains records with a single field:

• display_text: The complete Phi-3 formatted conversation

🎯 Intended Use Cases

Primary Applications

1. Edge Device Deployment: Run Phi-3 on industrial gateways or microcontrollers
2. Operator Assistance: Provide real-time explanations to maintenance personnel
3. Anomaly Root Cause Analysis: Go beyond detection to explanation
4. Multi-Domain Industrial AI: Single model for faults, security, and predictive maintenance

Suitable Models

• Phi-3 Mini (3.8B parameters) - Recommended
• TinyLlama (1.1B parameters)
• Qwen-2.5-Coder (1.5B parameters)
• Any SLM supporting chat templates

💻 Usage

Load with Hugging Face Datasets

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("your-username/Edge-Industrial-Anomaly-Phi3")

# Load individual splits
industrial = load_dataset("your-username/Edge-Industrial-Anomaly-Phi3", data_files="industrial_faults_phi3.jsonl")
edge_iiot = load_dataset("your-username/Edge-Industrial-Anomaly-Phi3", data_files="edge_iiot_phi3.jsonl")
predictive = load_dataset("your-username/Edge-Industrial-Anomaly-Phi3", data_files="predictive_maint_phi3.jsonl")

Fine-Tuning Example

from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")

# Load dataset
dataset = load_dataset("your-username/Edge-Industrial-Anomaly-Phi3")

# Configure training
training_args = TrainingArguments(
    output_dir="./phi3-industrial-anomaly",
    per_device_train_batch_size=4,
    num_train_epochs=3,
    learning_rate=2e-5,
    fp16=True,
    logging_steps=10,
)

# Initialize trainer
trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=dataset["train"],
    dataset_text_field="display_text",
)

# Train
trainer.train()

📊 Dataset Statistics

Dataset	Entries	Avg Length	Anomaly %
Industrial Faults	~10,000	180 tokens	~45%
Edge-IIoTset	~10,000	165 tokens	~35%
Predictive Maint	~10,000	175 tokens	~40%
Total	~30,000	173 tokens	~40%

⚙️ Data Processing Pipeline

The dataset was created using an automated curation pipeline:

1. Download: Kaggle API pulls raw CSV datasets
2. Normalization: Column names standardized, missing values handled
3. Reasoning Synthesis: Rule-based logic generates explanations:
   • Temperature > 75°C → "Critical overheating"
   • Vibration > 3.5 m/s → "Mechanical imbalance"
   • Pressure anomalies → "Outside operational range"
4. Format Conversion: Transform to Phi-3 chat template
5. Export: Save as JSONL for easy loading

🔬 Evaluation Metrics

For models trained on this dataset, consider evaluating:

1. Classification Accuracy: Can the model correctly identify NORMAL vs ANOMALY?
2. Reasoning Quality: Does the explanation match the sensor values?
3. Hallucination Rate: Does the model invent sensor readings not in the input?
4. Deployment Metrics: Inference latency on target edge hardware

📝 Limitations

• Synthetic Reasoning: The natural language explanations are rule-based, not human-annotated
• Domain Coverage: Limited to three specific industrial domains
• Sensor Diversity: Only includes common sensor types (temp, pressure, vibration, etc.)
• Language: English only

🤝 Citation

If you use this dataset, please cite the original Kaggle sources:

@misc{edge-industrial-anomaly-phi3,
  author = {Your Name},
  title = {Edge-Industrial-Anomaly-Phi3: A Curated Dataset for Small Language Models},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/your-username/Edge-Industrial-Anomaly-Phi3}}
}

📜 License

This dataset is released under the Apache 2.0 License, consistent with the source datasets and Phi-3 model licensing.

🛠️ Contributing

Found an issue or want to improve the dataset? Contributions are welcome:

• Report issues on GitHub
• Suggest additional industrial domains
• Propose improved reasoning templates

📧 Contact

For questions or collaboration opportunities, please open an issue on the dataset repository.

---

Built for the Edge AI Community 🚀