VertexAGI/cyberdata-2-medium

CyberData 2 Medium

25,000 chat examples for fine-tuning a model to write secure code, find vulnerabilities, reason about them, and work as a coding agent. Assembled entirely from existing public datasets; no model was run to write any of it.

---

The CyberData 2 family

Size	Repo	Examples	Train	Valid
Small	VertexAGI/cyberdata-2-small	15,000	14,050	950
Medium	VertexAGI/cyberdata-2-medium	25,000	23,451	1,549
Large	VertexAGI/cyberdata-2-large	65,000	60,812	4,188
Full	VertexAGI/cyberdata-2-full	174,877	164,357	10,520

CyberData 2 Medium is a nested subset of the larger sizes: each source contributes a fixed share (capped at what exists), always the first rows of a fixed order, so Small is inside Medium, Medium inside Large and Large inside Full, and every row keeps the same train/valid split in every size. Shares are set by goal (agentic coding about 29%, writing code and finding vulnerabilities about 37%, reasoning about 9%, supporting knowledge about 21%), not by how much of each source exists.

What this size teaches

Goal	Rows	Share
Writing secure code and detection rules (fix, rewrite, write-a-rule)	4,356	17.4%
Finding vulnerabilities (code review, weakness identification)	3,921	15.7%
Cyber reasoning (step-by-step vulnerability analysis with a verdict)	2,494	10.0%
Agentic coding and tool use (security and software-engineering agents)	6,489	26.0%
Security knowledge, Q&A and threat intelligence (supporting material)	7,740	31.0%

Where every row comes from (provenance)

Each row has a provenance tag in row_index.csv, so you can train on a stricter subset (for example, drop third-party-synthetic).

provenance	rows	share	what it means
authoritative-structured	10,752	43.0%	real records (MITRE entries, Sigma rules, KEV entries, real CVE-fix code) turned into chat rows by fixed templates; every fact is copied or composed from a field of the record
agent-trajectories	5,103	20.4%	security-agent and software-engineering agent trajectories
third-party-synthetic	2,771	11.1%	AI-written Q&A by the dataset authors; the least reliable rows, so they are FILTERED: any row citing a CWE id that does not exist, or naming a CWE or ATT&CK technique with a name that does not match MITRE's data, is dropped (on 3,000-row samples roughly 15-19% of rows that cite ids had a nonexistent or mismatched one), as are refusals and duplicate questions
reasoning-verified	2,494	10.0%	reasoning traces written by DeepSeek-R1-0528 for the OpenVul authors and kept only when the final verdict matched the ground-truth label
grounded-third-party	1,524	6.1%	instruction sets written by their authors from real sources, who state their claims are checked against the source
verified-executed	1,386	5.5%	agent steps whose episodes were executed against a local environment and passed a deterministic reward check
human-annotated	554	2.2%	expert-annotated threat-report text turned into extraction tasks by a template
nist-excerpt	416	1.7%	answers grounded in excerpts of NIST publications

Sources

Source	Rows	%	Provenance	License / terms	What it contributes
MITRE CWE	1,247	5.0%	authoritative-structured	MITRE CWE terms of use (royalty-free; keep the MITRE copyright notice)	Weakness definitions, mitigations, detection methods, and the entries' own bad/good code examples
MITRE CAPEC	277	1.1%	authoritative-structured	MITRE CAPEC terms of use (royalty-free; keep the MITRE copyright notice)	Attack-pattern descriptions, prerequisites, consequences, mitigations
MITRE ATT&CK (enterprise)	1,109	4.4%	authoritative-structured	MITRE ATT&CK terms of use (royalty-free; keep the MITRE copyright notice)	Techniques, mitigations, detection analytics, groups, software, campaigns
SigmaHQ detection rules	2,078	8.3%	authoritative-structured	Detection Rule License 1.1 (attribution to the rule authors, kept inside each rule)	Real rules: write a rule from a description, explain a rule
CISA Known Exploited Vulnerabilities	499	2.0%	authoritative-structured	US-government work, public domain	Exploited-in-the-wild status and required action per CVE
PrimeVul (paired vulnerable/fixed C/C++ functions)	2,356	9.4%	authoritative-structured	MIT (train + valid pairs only; the test pairs are never used)	Review a vulnerable function, review its patched version, fix it
CVEfixes (real CVE fix commits, many languages)	3,186	12.7%	authoritative-structured	Apache-2.0 per the Hub listing (the original CVEfixes release is CC BY 4.0)	The same three code tasks across PHP/Python/JavaScript/Java/Go/Ruby/C/C++ and others
OpenVul rejection-sampling vulnerability reasoning (Leopo1d)	2,494	10.0%	reasoning-verified	Apache-2.0 (reasoning distilled from DeepSeek-R1-0528 by the dataset's authors)	Step-by-step reasoning about whether C/C++ code is vulnerable, kept only when the verdict matched the label; vulnerable and patched versions in pairs
Annotated threat-report corpus	554	2.2%	human-annotated	CC-BY-4.0 per the Hub tag (the dataset card has no text)	Entity extraction and relationship listing from real threat-report excerpts
CTI instruction dataset (reloading0101)	1,524	6.1%	grounded-third-party	CC-BY-4.0	CVE/ATT&CK/KEV/Sigma-grounded threat-intelligence instructions and follow-ups; its authors state every claim is checked against the source
NIST SP excerpt instructions (nuhmanpk)	416	1.7%	nist-excerpt	CC-BY-4.0 (source documents are NIST publications)	Answers grounded in excerpts of NIST SP 800-series documents
OWASP defensive repair trajectories (Humanlearning)	1,386	5.5%	verified-executed	Apache-2.0	Agent steps from episodes that were executed and reward-verified in a local environment (successful episodes only, no anti-cheat flags)
Trendyol cybersecurity instructions	831	3.3%	third-party-synthetic	Apache-2.0	AI-written defensive Q&A, filtered (see below)
Fenrir v2.1	554	2.2%	third-party-synthetic	Apache-2.0	AI-written defensive Q&A, filtered, de-duplicated against Trendyol
CyberNative code vulnerability pairs	1,109	4.4%	third-party-synthetic	Apache-2.0	Review vulnerable code and write the safer version
YARA / Suricata rules Q&A	277	1.1%	third-party-synthetic	Apache-2.0	Write a YARA or Suricata rule
0xKitkat/AgentForge-1152 (security split)	277	1.1%	agent-trajectories	Apache-2.0	Defensive security-agent trajectories with tool use
nvidia/SWE-Hero-openhands-trajectories	1,639	6.6%	agent-trajectories	CC-BY-4.0	Software-engineering agent trajectories (OpenHands), length-capped at 120,000 characters
nvidia/Open-SWE-Traces (resolved runs only)	3,187	12.7%	agent-trajectories	CC-BY-4.0	Coding-agent trajectories whose patch was verified to fix the task (OpenHands / SWE-agent / mini-swe-agent), length-capped at 120,000 characters

Considered and left out

• trendmicro-ailab/Primus-Reasoning and Primus-Instruct: gated (access request required).
• rmems/incident-response-oncall-trajectories: tagged Apache-2.0, but its own rights note says it is not training data for any model-weight update.
• netgoat-ai/Koda-IDS-CyberReasoning: inspected and rejected. After quality filtering only templated scenarios for one vendor's tool API remained.
• ayshajavd/code-security-vulnerability-dataset: mostly a repackaging of BigVul (no clear license), PrimeVul-derived samples and CyberNative, so it added little that was new.
• Benchmark sets (CyberMetric, MMLU computer security, CTI-Bench): kept out so they stay usable for evaluation; a scan found none of their questions verbatim here.

Splits and checks

train.jsonl has 23,451 rows and valid.jsonl has 1,549. Rows derived from the same object (a CWE entry, a Sigma rule, a code-fix pair, an agent episode, a vulnerable/patched reasoning pair, a document chunk) always share one split. Checks run before release (build_scripts/verify_sizes.py): exact row counts; unique ids; strict nesting between sizes; the same split for every row in every size; every line parses with valid roles and an assistant last turn; no row mentions any of the 916 CVEs we hold out for a CVE-recall evaluation; no verbatim CyberMetric / MMLU-security question appears.

Format

Each line is {"messages": [...]}, standard chat SFT format. Agent rows carry tool_calls; reasoning rows keep the source's <think>...</think> block followed by the final answer. row_index.csv has id, source, task, provenance, split, group per row.

from datasets import load_dataset
ds = load_dataset("VertexAGI/cyberdata-2-medium")
print(ds["train"][0]["messages"])

Limitations

• Vulnerability finding is supply-limited. The real vulnerable-and-fixed code comes from PrimeVul, CVEfixes and the OpenVul reasoning set; there is no more openly licensed labeled data we trusted. Vulnerability detection is a hard task: training on these labels does not by itself guarantee good detection.
• Reasoning covers one task. The only verified reasoning set we found is vulnerability detection in C/C++. There is no verified reasoning data here for incident response or threat analysis.
• Agent trajectories are length-capped at 120,000 characters (about 30k tokens), so the longest runs are absent.
• Templated rows teach the answer format and the content of the records as written; they are not a verified knowledge base. In our tests none of the 8B base models we tried could recall facts about CVEs they had not seen, so for CVE-specific facts use lookup, not memory.
• PrimeVul and CVEfixes CWE labels come from NVD and are sometimes loose. The third-party-synthetic and grounded-third-party rows can still contain errors beyond the ids we can check.

Licensing and attribution

A mixture of independently licensed sources (see the table). Required attribution: (c) The MITRE Corporation for CWE, CAPEC and ATT&CK content; SigmaHQ rule authors (Detection Rule License 1.1; author fields are kept in the rule text); CC-BY-4.0 sources (nvidia SWE-Hero and Open-SWE-Traces, reloading0101, nuhmanpk, mrmoor) must be credited as listed. The OpenVul reasoning was written by DeepSeek-R1-0528; check DeepSeek's terms for your use. No source is gated or carries a non-commercial restriction.