CVE-TO-MITRE

	---
	license: mit
	task_categories:
	- text-classification
	- zero-shot-classification
	language:
	- en
	size_categories:
	- 10K<n<100K
	tags:
	- security
	- mitre-attack
	- cve
	- cybersecurity
	- vulnerability
	- threat-intelligence
	pretty_name: MITRE CVE Community Pack
	---

	# MITRE CVE Community Pack

	A comprehensive dataset for training multi-label classifiers to predict MITRE ATT&CK techniques from vulnerability descriptions.

	## Dataset Description

	This dataset contains vulnerability descriptions paired with MITRE ATT&CK technique labels, designed for training and evaluating multi-label text classification models in the cybersecurity domain.

	### Key Features

	- Dual-source data: 10,000 real CVE descriptions and 10,000 synthetic vulnerability descriptions
	- Multi-label classification: Each vulnerability can be associated with multiple MITRE ATT&CK techniques
	- Cybersecurity-focused: Specifically curated for security research and threat intelligence applications
	- Ready for training: Compatible with Hugging Face Datasets and Transformers libraries

	## Dataset Structure

	### Data Files

	- `real_10k.jsonl`: 10,000 real vulnerability descriptions from CVE (Common Vulnerabilities and Exposures)
	- `synth_10k.jsonl`: 10,000 synthetically generated vulnerability descriptions
	- `attaquemitre2.csv`: MITRE ATT&CK technique reference (ID, name, description, platforms)

	### Data Format

	Each line in the JSONL files contains:

	```json
	{
	"text": "Description of the vulnerability...",
	"labels": ["T1078", "T1203", "T1059"]
	}
	```

	- `text`: Natural language description of the vulnerability
	- `labels`: List of MITRE ATT&CK technique IDs (multi-label)

	### Label Space

	The dataset uses MITRE ATT&CK technique IDs as labels (e.g., T1078, T1203, T1059). The `attaquemitre2.csv` file provides detailed information about each technique including:
	- Technique ID
	- Technique name
	- Description
	- Target platforms

	## Usage

	### Loading the Dataset

	```python
	import json
	from datasets import Dataset

	# Load real CVE data
	real_data = []
	with open('real_10k.jsonl', 'r', encoding='utf-8') as f:
	for line in f:
	real_data.append(json.loads(line))

	# Load synthetic data
	synth_data = []
	with open('synth_10k.jsonl', 'r', encoding='utf-8') as f:
	for line in f:
	synth_data.append(json.loads(line))

	# Combine datasets
	all_data = real_data + synth_data
	dataset = Dataset.from_list(all_data)
	```

	### Training Example (ModernBERT)

	```python
	import numpy as np
	from sklearn.model_selection import train_test_split
	from transformers import AutoTokenizer, AutoModelForSequenceClassification
	from transformers import Trainer, TrainingArguments, DataCollatorWithPadding
	from datasets import Dataset

	# Load data
	texts = [item['text'] for item in all_data]
	all_labels = sorted(list(set(l for item in all_data for l in item['labels'])))
	label2id = {l: i for i, l in enumerate(all_labels)}
	id2label = {i: l for i, l in enumerate(all_labels)}

	# Create multi-hot encoded labels
	labels = np.zeros((len(texts), len(all_labels)), dtype=np.float32)
	for i, item in enumerate(all_data):
	for lbl in item['labels']:
	labels[i, label2id[lbl]] = 1.0

	# Split data
	train_texts, val_texts, train_labels, val_labels = train_test_split(
	texts, labels, test_size=0.1, random_state=42
	)

	# Tokenize
	tokenizer = AutoTokenizer.from_pretrained("answerdotai/ModernBERT-large")
	train_ds = Dataset.from_dict({"text": train_texts, "labels": train_labels}).map(
	lambda x: tokenizer(x["text"], truncation=True, max_length=1024),
	batched=True
	)

	# Model setup
	model = AutoModelForSequenceClassification.from_pretrained(
	"answerdotai/ModernBERT-large",
	num_labels=len(id2label),
	problem_type="multi_label_classification",
	id2label=id2label,
	label2id=label2id
	)
	```

	## Dataset Statistics

	- Total samples: 20,000 (10,000 real + 10,000 synthetic)
	- Average labels per sample: ~1.2 techniques
	- Unique techniques: Covers a wide range of MITRE ATT&CK tactics
	- Language: English
	- Domain: Cybersecurity / Vulnerability Assessment

	## Applications

	- Vulnerability analysis: Automatically categorize CVEs by ATT&CK techniques
	- Threat intelligence: Enhance CTI platforms with automated labeling
	- Security research: Train models for technique prediction
	- Incident response: Quickly identify potential attack techniques in vulnerabilities
	- Risk assessment: Prioritize vulnerabilities based on associated techniques

	## Considerations

	- Multi-label nature: Each vulnerability may map to multiple ATT&CK techniques
	- Class imbalance: Some techniques are more common than others
	- Synthetic data: The `synth_10k.jsonl` file contains AI-generated descriptions for training augmentation
	- Label granularity: Uses technique-level labels (e.g., T1078) rather than tactic or sub-technique

	## Reference

	- MITRE ATT&CK Framework: https://attack.mitre.org/
	- CVE Database: https://cve.mitre.org/

	## Citation

	If you use this dataset, please cite:
	```
	@dataset{mitre_cve_community_pack,
	title={MITRE CVE Community Pack},
	author={Community},
	year={2024},
	url={https://huggingface.co/datasets/[your-username]/MITRE_CVE_Community_Pack}
	}
	```

	## License

	MIT License

	---

	Note: This dataset is designed for research and educational purposes in cybersecurity. Always validate predictions with human experts in production security environments.