VertexAGI/cyberdata-small

CyberData Small

A 15,000-example SFT dataset combining verified agentic-coding trajectories, cybersecurity agent behavior, and structured vulnerability intelligence — built entirely from non-gated, redistributable sources.

---

The CyberData family

CyberData Small (this dataset, originally released as CyberData 1 / VertexAGI/cyberdata-1, the old name redirects here) is the smallest of four nested sizes:

Size	Repo	Examples	Train	Valid
Small	VertexAGI/cyberdata-small	15,000	13,800	1,200
Medium	VertexAGI/cyberdata-medium	25,000	23,009	1,991
Large	VertexAGI/cyberdata-large	65,000	59,745	5,255
Full	VertexAGI/cyberdata-full	479,214	440,873	38,341

Every Small example is also in Medium, Large and Full, and Small's validation examples remain validation examples in all of them.

What this is

CyberData Small is a curated mixture of five public Hugging Face datasets, resampled and templated into one consistent messages-format SFT dataset. It skews heavily toward real, execution-verified agentic coding (tool use, multi-turn trajectories, actual resolved software-engineering tasks), with a smaller layer of dedicated cybersecurity agent behavior and structured vulnerability knowledge (CVE/CWE, OWASP, MITRE ATT&CK).

Composition

Source	Examples	%	What it contributes
nvidia/SWE-Hero-openhands-trajectories	6,000	40.0%	Execution-based software-engineering agent trajectories (OpenHands framework)
nvidia/Open-SWE-Traces	5,339	35.6%	Multi-harness (OpenHands, SWE-agent, mini-swe-agent) coding-agent trajectories, filtered to resolved=1 only — the agent's patch was verified to actually fix the task
0xKitkat/AgentForge-1152 (security split)	576	3.8%	Evidence-grounded defensive-cybersecurity agent trajectories: tool use, failed-check recovery, evidence-based findings
ismailtasdelen/unified-vulnerability-intelligence-dataset	500	3.3%	Structured vulnerability knowledge (CWE, CAPEC, MITRE ATT&CK, OWASP, CVSS, remediation, detection)
stasvinokur/cve-and-cwe-dataset-1999-2025	2,585	17.2%	Real NVD CVE records (1999–2025) with severity, CVSS, and CWE classification
Total	15,000	100%

Why Open-SWE-Traces is short of its original 5,500 target

The resolved=1 filter (only agent trajectories whose patch was verified to fix the task) removes the large majority of rows in some harness/teacher-model subsets. After sampling across all 13 harness/teacher/source-dataset combinations in the upstream repo, only 5,339 genuinely resolved trajectories were available within the sampled shards. The remaining 161 examples needed to reach 15,000 total were filled from stasvinokur/cve-and-cwe-dataset-1999-2025, which has ample supply (280,700 rows) and was already in the mix.

What was deliberately left out

ethanolivertroy/nist-cybersecurity-training was evaluated and excluded. Its published schema (id/text/embedding/metadata) does not match its own README's description of a clean system/user/assistant messages format — the actual text column is a flat blob with no separable ground-truth answer. Rather than have another model invent an answer to pair with the question embedded in that blob, it was dropped entirely.

Templating methodology (UVID and CVE/CWE)

nvidia/SWE-Hero-openhands-trajectories, nvidia/Open-SWE-Traces, and 0xKitkat/AgentForge-1152 were already in multi-turn messages format and are used as-is (each trajectory keeps its own source-specific system prompt describing its tools and environment).

ismailtasdelen/unified-vulnerability-intelligence-dataset and stasvinokur/cve-and-cwe-dataset-1999-2025 are structured metadata tables, not pre-written conversations. Each was converted into a single-turn Q&A pair using a deterministic template: every fact stated in the answer comes from a column already present in that row (category, severity, CWE, CVSS score, remediation, detection guidance, CVE description, etc.). A field left blank in the source is simply omitted from the answer — nothing is inferred, estimated, or generated by another model to fill a gap.

Format

Each row is {"messages": [...]}, standard chat SFT format. Some rows additionally carry tool_calls on assistant turns (the agentic-coding trajectories).

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

• train.jsonl — 13,800 examples
• valid.jsonl — 1,200 examples (held out, not used for template/dedup verification — a genuine 8% split)

Licensing

This is a mixture of five independently-licensed sources: CC-BY-4.0 (nvidia/SWE-Hero-openhands-trajectories, nvidia/Open-SWE-Traces), Apache-2.0 (0xKitkat/AgentForge-1152), MIT (ismailtasdelen/unified-vulnerability-intelligence-dataset), and CC0-1.0 (stasvinokur/cve-and-cwe-dataset-1999-2025). No source in this mix is gated or carries a non-commercial restriction. If you redistribute this dataset, retain attribution to each upstream source above, per their respective licenses (required for the CC-BY-4.0 portions).

Quality checks performed

• Zero duplicate example IDs across the full 15,000.
• Zero malformed rows (every row has ≥2 messages with valid role/content).
• Open-SWE-Traces rows are restricted to resolved=1 — the agent's patch was independently verified against the task's test suite.
• Every row from every source was verified against the real, live Hugging Face repo (row counts, schema, license) before inclusion — not taken from a dataset card's stated numbers alone, several of which turned out to be inaccurate on inspection.

Limitations

This is a resampled mixture, not a from-scratch curated dataset — quality is bounded by the upstream sources. The CVE/CWE and UVID portions are single-turn Q&A synthesized from structured data via template, not natural human-written conversation. The agentic-coding portion dominates the mix (75.6%); the cybersecurity-agent and vulnerability-knowledge portions are comparatively small (23.4% combined).