fmt0816/soc-triage-dataset

SOC Triage Dataset

A synthetic dataset of security alerts with expert-level triage decisions for training security operations models.

Dataset Description

Summary

This dataset contains synthetic security alerts and corresponding triage recommendations, designed for training and evaluating machine learning models for Security Operations Center (SOC) automation.

Each sample includes:

• Detailed security alert with indicators of compromise
• User context (department, role, VIP status)
• Asset context (criticality, data classification)
• Environmental context (business hours, threat level)
• Expert triage decision with reasoning

Supported Tasks

• Security Alert Classification: Classify alerts into triage decisions
• Priority Prediction: Predict appropriate priority levels
• Text Generation: Generate detailed triage recommendations

Languages

English (en)

Dataset Structure

Data Instances

Each instance contains a conversation in chat format:

{
  "messages": [
    {
      "role": "system",
      "content": "You are an expert Security Operations Center (SOC) analyst..."
    },
    {
      "role": "user", 
      "content": "Analyze the following security alert...\n\n## Alert Details\n..."
    },
    {
      "role": "assistant",
      "content": "## Triage Recommendation\n\n### Decision Summary\n| Field | Value |\n..."
    }
  ]
}

With metadata (when include_metadata=True):

{
  "messages": [...],
  "_metadata": {
    "alert": {
      "alert_id": "550e8400-e29b-41d4-a716-446655440000",
      "timestamp": "2024-01-15T14:32:00",
      "source_system": "CrowdStrike Falcon",
      "category": "malware",
      "severity": "high",
      "title": "Suspicious executable detected on endpoint",
      "description": "...",
      "indicators": {...},
      "user_context": {...},
      "asset_context": {...},
      "environment_context": {...}
    },
    "triage": {
      "decision": "escalate",
      "priority": 2,
      "confidence_score": 0.92,
      "reasoning": "...",
      "recommended_actions": [...],
      "escalation_required": true,
      "escalation_target": "Incident Response Team"
    }
  }
}

Data Fields

Alert Fields

Field	Type	Description
alert_id	string	Unique identifier
timestamp	string	ISO 8601 timestamp
source_system	string	Detection system name
category	string	Alert category (12 types)
severity	string	critical/high/medium/low/informational
title	string	Alert title
description	string	Detailed description
affected_assets	list[string]	Affected hostnames
indicators	dict	Category-specific IOCs
user_context	dict	User information
asset_context	dict	Asset information
environment_context	dict	Environmental factors
raw_log	string	Simulated raw log entry
mitre_techniques	list[string]	MITRE ATT&CK IDs

Triage Fields

Field	Type	Description
decision	string	escalate/investigate/monitor/false_positive/close
priority	int	1 (highest) to 5 (lowest)
confidence_score	float	0.0 to 1.0
reasoning	string	Explanation of decision
key_factors	list[string]	Key decision factors
recommended_actions	list[string]	Remediation steps
escalation_required	bool	Whether to escalate
escalation_target	string	Escalation destination
estimated_impact	string	none/low/moderate/high/severe
estimated_urgency	string	immediate/hours/day/week
additional_investigation	list[string]	Further investigation steps
ioc_extraction	list[string]	Extracted IOCs

Alert Categories

Category	Description	Examples
malware	Malware detection	Ransomware, trojans, cryptominers
phishing	Email threats	Credential harvesting, BEC
brute_force	Auth attacks	Password spraying, credential stuffing
data_exfiltration	Data theft	Unauthorized transfers, USB copying
privilege_escalation	Privilege abuse	Token manipulation, UAC bypass
lateral_movement	Network spread	Pass-the-hash, WMI execution
command_and_control	C2 activity	Beaconing, DNS tunneling
insider_threat	User risk	Anomalous behavior, data hoarding
policy_violation	Compliance	Unauthorized software, config drift
vulnerability_exploit	CVE attacks	SQL injection, RCE attempts
reconnaissance	Scanning	Port scans, enumeration
denial_of_service	DDoS	Volumetric attacks, slowloris

Data Splits

Split	Samples	Purpose
train	8,000	Model training
validation	1,000	Hyperparameter tuning
test	1,000	Final evaluation

Dataset Creation

Generation Process

1. Template-Based Generation: Realistic alert titles and descriptions from security expert templates
2. Indicator Generation: Category-specific IOCs (hashes, IPs, CVEs, etc.)
3. Context Generation: Randomized but coherent user/asset/environment contexts
4. Triage Logic: Rule-based expert decisions following security best practices

Triage Decision Logic

Decisions are determined by:

• Escalate: Critical severity, successful attacks, lateral movement, threat intel matches
• Investigate: High severity, suspicious indicators, context anomalies
• Monitor: Low severity, partial indicators, uncertain outcomes
• False Positive: Change window activity, known authorized tools
• Close: Informational alerts, no action needed

Priority is influenced by:

• Severity weight (critical=1 to info=5)
• VIP user involvement (-1 priority)
• Critical asset involvement (-1 priority)
• Elevated threat level (-1 priority)

Source Data

All data is synthetically generated. No real security incidents or personal data.

Considerations

Biases

• Balanced across alert categories by default
• Real-world distributions may differ significantly
• Some categories have more varied templates than others

Limitations

• Synthetic data may not capture all real-world edge cases
• Triage logic is rule-based, not from actual analyst decisions
• Context combinations may occasionally be unrealistic
• English language only

Recommendations

• Supplement with real (anonymized) alert data if available
• Adjust decision thresholds based on organizational risk tolerance
• Use for initial training, then fine-tune with production feedback
• Always maintain human oversight for critical decisions

Usage

Load Dataset

from datasets import load_dataset

# From Hugging Face Hub
dataset = load_dataset("fmt0816/soc-triage-dataset")

# Access splits
train_data = dataset["train"]
val_data = dataset["validation"]
test_data = dataset["test"]

Generate Custom Dataset

from soc_triage_agent import SecurityAlertGenerator

generator = SecurityAlertGenerator(seed=42)
dataset = generator.generate_dataset(
    num_samples=10000,
    format_type="chat",
    balanced=True,
    include_metadata=True
)

Format Variants

# OpenAI chat format (default)
generator.generate_dataset(format_type="chat")

# ShareGPT format
generator.generate_dataset(format_type="sharegpt")

# Instruction format
generator.generate_dataset(format_type="instruction")

# Hugging Face format
generator.generate_dataset(format_type="huggingface")

Citation

@dataset{soc_triage_dataset,
  title = {SOC Triage Dataset: Synthetic Security Alerts for ML Training},
  author = {ftrout},
  year = {2025},
  url = {https://huggingface.co/datasets/fmt0816/soc-triage-dataset},
  license = {Apache-2.0}
}

License

Apache License 2.0