alucent/mirror-cve-to-cwe

CVE-to-CWE + ATT&CK Classification Dataset

A dataset for mapping CVE (Common Vulnerabilities and Exposures) descriptions to CWE (Common Weakness Enumeration) categories and MITRE ATT&CK techniques. Built from the National Vulnerability Database (NVD) with AI-assisted label refinement.

Tasks

1. CVE → CWE classification (single-label, 205 classes) — map vulnerability description to weakness type
2. CVE → ATT&CK technique mapping (multi-label, 361 techniques) — map vulnerability to attack techniques

Dataset Summary

Split	Samples	CWE Label Source	ATT&CK Coverage
Train	234,770	Claude Sonnet 4.6 relabeled	97.2%
Validation	27,896	Agreement-filtered (NVD == Sonnet)	98.2%
Test	27,780	Agreement-filtered (NVD == Sonnet)	98.2%

• CWE classes: 205
• ATT&CK techniques: 361 unique (multi-label per sample)
• Years covered: 1999–2026
• Source: NVD (National Vulnerability Database)

Data Fields

Field	Type	Description
cve_id	string	CVE identifier (e.g., "CVE-2024-12345")
description	string	Vulnerability description from NVD
cwe_id	string	CWE category (e.g., "CWE-79")
label	int	Numeric CWE label ID (0–204), see label_map.json
attack_techniques	list[string]	MITRE ATT&CK technique IDs (e.g., ["T1190", "T1059.007"])

CWE Label Quality

The original NVD CWE labels are known to be noisy (often too generic, e.g., CWE-20 "Improper Input Validation" used as catch-all). To improve label quality:

1. All 318,979 CVE descriptions were relabeled using Claude Sonnet 4.6 via the Anthropic Batch API (~$395 total cost)
2. 73.1% exact CWE ID agreement between NVD and Sonnet labels (84.5% with hierarchy-aware matching, indicating ~half of disagreements are granularity differences)
3. Validation and test sets contain only agreement-filtered samples where NVD and Sonnet labels match exactly
4. Training set uses Sonnet labels for all samples (including disagreements and previously unlabeled CVEs)

This means val/test are a high-confidence subset where two independent labelers agree, but biased toward unambiguous cases — samples where labelers disagree are excluded.

Manual validation: A random sample of 100 NVD-Sonnet disagreements was manually reviewed — Sonnet was clearly more accurate in 72% of cases, NVD in only 3%, with the remainder being ambiguous or both-acceptable (hierarchy/sibling CWEs).

Top 20 CWE Classes

CWE	Name	Train Count	%
CWE-79	Cross-site Scripting	33,858	14.4%
CWE-89	SQL Injection	15,619	6.7%
CWE-22	Path Traversal	8,047	3.4%
CWE-121	Stack-based Buffer Overflow	7,651	3.3%
CWE-862	Missing Authorization	7,533	3.2%
CWE-78	OS Command Injection	7,132	3.0%
CWE-125	Out-of-bounds Read	6,770	2.9%
CWE-200	Information Exposure	6,516	2.8%
CWE-787	Out-of-bounds Write	6,508	2.8%
CWE-20	Improper Input Validation	6,299	2.7%
CWE-352	CSRF	6,270	2.7%
CWE-416	Use After Free	6,009	2.6%
CWE-119	Buffer Overflow	5,943	2.5%
CWE-400	Resource Exhaustion	5,809	2.5%
CWE-284	Improper Access Control	5,270	2.2%
CWE-476	NULL Pointer Dereference	4,931	2.1%
CWE-122	Heap-based Buffer Overflow	4,787	2.0%
CWE-434	Unrestricted Upload	3,697	1.6%
CWE-306	Missing Authentication	3,313	1.4%
CWE-190	Integer Overflow	3,210	1.4%

Top 15 ATT&CK Techniques

Technique	Name	Count
T1190	Exploit Public-Facing Application	127,837
T1203	Exploitation for Client Execution	45,480
T1499	Endpoint Denial of Service	30,344
T1068	Exploitation for Privilege Escalation	25,481
T1059.007	JavaScript	24,185
T1059	Command and Scripting Interpreter	19,199
T1005	Data from Local System	18,887
T1552	Unsecured Credentials	10,495
T1078	Valid Accounts	6,285
T1557	Adversary-in-the-Middle	4,667
T1189	Drive-by Compromise	3,955
T1110	Brute Force	2,249
T1083	File and Directory Discovery	2,076
T1210	Exploitation of Remote Services	1,805
T1040	Network Sniffing	1,547

Data Decontamination

All 2,000 CVEs from the CTI-Bench benchmark (NeurIPS 2024) have been removed from all splits to enable clean external evaluation.

Trained Models

• xamxte/cwe-classifier-roberta-base — RoBERTa-base fine-tuned on this dataset. 87.4% top-1 accuracy on agreement-filtered test set, competitive with best open-weight models on CTI-Bench RCM (75.6% strict).

Usage

from datasets import load_dataset

dataset = load_dataset("xamxte/cve-to-cwe")

# Access splits
train = dataset["train"]
val = dataset["validation"]
test = dataset["test"]

print(train[0])
# {
#   'cve_id': 'CVE-2025-7782',
#   'description': 'The WP JobHunt plugin for WordPress...',
#   'cwe_id': 'CWE-862',
#   'label': 186,
#   'attack_techniques': ['T1190', 'T1059.007']
# }

CWE Hierarchy Note

This dataset uses specific (child) CWE categories where possible, rather than generic parent categories. For example, buffer overflow vulnerabilities are labeled as CWE-121 (Stack Buffer Overflow) or CWE-122 (Heap Buffer Overflow) rather than the generic CWE-119 (Buffer Overflow). This provides more actionable information for vulnerability triage.

Limitations

• Single-label CWE: Each CVE is assigned exactly one CWE, though some vulnerabilities may involve multiple weakness types
• Description-only: Classification is based solely on the text description; CVSS scores, CPE data, and other metadata are not included
• English only: All descriptions are in English (NVD standard)

Paper

📄 Fine-tuning RoBERTa for CVE-to-CWE Classification: A 125M Parameter Model Competitive with LLMs

Citation

If you use this dataset, please cite:

@article{mosievskiy2026cwe,
  title={Fine-tuning RoBERTa for CVE-to-CWE Classification: A 125M Parameter Model Competitive with LLMs},
  author={Mosievskiy, Nikita},
  journal={arXiv preprint arXiv:2603.14911},
  year={2026}

License

CC-BY-4.0