ronaldocloud/cyberusecase-v1.0

Cybersecurity SOC Fine-Tuning Dataset — 17.5k Real CVEs (2018–2026) + SOC Knowledge

A large supervised fine-tuning (SFT) dataset for teaching an LLM expert-level cybersecurity reasoning across vulnerability management, SOC alert triage, detection engineering, threat intelligence & hunting, incident response, and cloud/DevSecOps.

It combines 17,590 real CVEs (2018–2026) pulled from the NIST NVD data feeds with a hand-curated set of 65 landmark CVEs (rich, multi-angle coverage) and a body of practical SOC knowledge. Records use the HuggingFace chat (messages) format, ready for TRL SFTTrainer, Axolotl, Unsloth, and LLaMA-Factory.

Dataset at a glance

Total examples	59,636
Train / Validation	57,847 / 1,789 (~97/3)
Format	JSONL, OpenAI-style messages (system/user/assistant)
Real CVEs (NVD 2018–2026)	17,590 (99% with CVSS, 81% with mapped CWE)
Curated landmark CVEs	65 (13 task framings each)
Unique CVE IDs referenced	~17,900
System-prompt variety	15 expert personas + ~12% with no system prompt
Short/terse-query examples	included (users often type just a CVE ID)
Language	English
License	CC-BY-4.0

Why it's built this way

• Breadth (real CVEs). ~17.5k genuine CVEs from NVD so the model sees the actual distribution of vulnerabilities, vendors, products, CVSS vectors and CWE classes from 2018 through 2026 — not a toy sample.
• Depth (curated CVEs). 65 landmark CVEs (Log4Shell, ProxyLogon/Shell, MOVEit, Citrix Bleed 1 & 2, PAN-OS, Ivanti, XZ backdoor, SharePoint ToolShell, Cisco SD-WAN, Chrome V8 zero-day, …) expanded into 13 task framings each with real exploitation, threat-actor, and detection detail.
• Robustness (prompt & query variety). 15 different system-prompt personas plus a large share of examples with no system prompt, and many short/terse user queries (CVE-2024-1234, log4shell fix, T1059?) so the fine-tuned model answers accurately whether the user is verbose or types two words.

Record schema

Each line is a JSON object. Some records include a system message; ~12% intentionally omit it.

{
  "messages": [
    {"role": "system", "content": "You are a detection engineer. ..."},
    {"role": "user", "content": "How do I detect exploitation of CVE-2023-34362?"},
    {"role": "assistant", "content": "Detecting exploitation of CVE-2023-34362 ..."}
  ],
  "category": "cve_bulk_detection"
}

The category field is metadata for filtering/analysis; drop it if your trainer expects only messages.

Load & fine-tune

from datasets import load_dataset
ds = load_dataset("<your-username>/cyber-soc-llm-dataset")
print(ds)  # train: 57847, validation: 1789

from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model="your-base-model",
    train_dataset=ds["train"],
    eval_dataset=ds["validation"],
    args=SFTConfig(max_seq_length=2048, packing=True),
)
trainer.train()

Most chat-SFT trainers apply the base model's chat template to messages automatically.

Content coverage

Real CVEs (bulk, from NVD 2018–2026)

Each real CVE is expanded into an overview, a remediation plan, and a rotating third framing (CVSS interpretation, CWE root-cause, or SOC detection), with ~1/3 also getting a terse short-query example. Remediation and detection are enriched by a CWE knowledge map (47 common CWEs → class explanation, prevention, and detection guidance), so answers are substantive rather than "just patch it".

Year distribution of the real-CVE sample (evenly drawn across each year):

Year	CVEs	Year	CVEs
2018	1,991	2023	2,155
2019	2,016	2024	2,129
2020	1,597	2025	1,849
2021	1,924	2026	1,641
2022	2,288

Curated landmark CVEs (13 framings each)

cve_overview, cve_cvss, cve_cwe, cve_exploitation, cve_remediation, cve_detection, cve_mitre, cve_triage, cve_ir, cve_threatintel, cve_quickfacts, cve_exec_summary, cve_prevention.

SOC knowledge

• Detection & triage — 20 MITRE ATT&CK techniques, Splunk SPL / Microsoft Sentinel KQL / Sigma / Falco detections, analyst triage workflows.
• Incident response — NIST SP 800-61 / SANS PICERL playbooks (ransomware, BEC, webshell, insider, cloud credential compromise).
• Threat intel & hunting — Pyramid of Pain, Diamond Model, Cyber Kill Chain, IOC vs IOA, hunt hypotheses, CTI tiers, purple teaming.
• Cloud / DevSecOps — S3 & IAM hardening, Kubernetes/container security, CI/CD secrets, IaC scanning, supply-chain defense, Zero Trust.

Category counts

Category	Count
cve_bulk_overview	17,590
cve_bulk_remediation	17,590
cve_bulk_detection	7,643
cve_bulk_short (terse queries)	5,864
cve_bulk_cvss	5,769
cve_bulk_cwe	4,178
curated cve_* (13 types × 65)	845
MITRE technique / detection / mitigation	60
threat_intel	23
cloud_devsecops	25
soc_triage / incident_response	25
detection_engineering / detection_concept	24

Sourcing & accuracy

Real CVE fields (description, CVSS score/vector/severity, CWE, affected vendor/product, publication date, reference URL) are extracted from the NIST NVD data feeds (via the community fkie-cad/nvd-json-data-feeds mirror). Curated landmark CVEs add exploitation context, associated threat actors, and ATT&CK mappings from public vendor and CISA advisories. Entries for 2025–2026 should be re-verified against primary sources before operational use, as recent-CVE details can be revised.

Detection queries and playbooks are illustrative reference implementations — tune all queries and thresholds to your environment before deployment.

Intended use

Supervised fine-tuning / instruction-tuning of LLMs for defensive cybersecurity assistance: SOC copilots, vulnerability triage, detection engineering support, IR guidance, and security education.

Limitations & responsible use

• Defensive focus. Written for defenders; no working exploit code or weaponization instructions.
• Templated breadth. Bulk CVE answers are generated from real NVD fields via templates, so phrasing repeats across CVEs (varied prompts/personas mitigate overfitting). The curated set and SOC banks provide depth and stylistic variety.
• Point-in-time. Refresh CVE/TTP content periodically; always validate against NVD/CISA/vendor primary sources.
• Not a source of truth. A fine-tuned model can still hallucinate — keep a human in the loop and verify CVE facts before acting.

Reproduction

The dataset is generated by scripts that pull the NVD mirror, build a CWE knowledge map, and template examples with system-prompt/query variety, then dedup and split. Regenerate to refresh with newer CVEs or to change the sampling size.

Citation

@misc{cyber_soc_llm_dataset,
  title  = {Cybersecurity SOC Fine-Tuning Dataset (17.5k Real CVEs 2018-2026 + SOC Knowledge)},
  year   = {2026},
  note   = {Instruction dataset for defensive security LLM fine-tuning}
}

License

Released under CC-BY-4.0. CVE data derives from public sources (NIST NVD, CISA, vendor advisories). You are responsible for compliance and for validating accuracy before operational use.