BlackGuards

BlackGuards is a cybersecurity-focused QLoRA adapter for GLM-4.7-Flash. It is intended for authorized security assessment, defensive engineering, incident response, threat intelligence, and malware analysis in isolated environments.

Status

This card is completed with the exact adapter revision, training counts, hardware time, loss curves, and evaluation results before weights are released. An adapter is not considered released while any placeholder remains.

Intended use

• CVE/CWE triage and remediation planning
• Detection engineering and incident-response playbooks
• Threat-intelligence and malware-behavior analysis
• Authorized red-team and purple-team exercise planning

Out-of-scope use

Unauthorized access, credential theft, malware deployment, destructive operations, evasion against third-party systems, or decisions without qualified human review.

Training

• Base: zai-org/GLM-4.7-Flash
• Method: 4-bit QLoRA SFT using MS-SWIFT
• Hardware: NVIDIA H100 80GB HBM3
• GPU count: 1
• Dataset manifest: included with this release
• Prepared rows: 396,081 train / 4,058 validation / 4,020 test
• Base revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
• Adapter revision: PENDING

Evaluation

PENDING: publish measured quality, grounding, regression, and safety results.

Limitations

The model can hallucinate vulnerability details, patches, indicators, and attribution. CVE and KEV status changes over time. Always verify against current NVD, CISA, vendor advisories, and organizational policy. This adapter does not provide authorization to test any system.

Attribution

The dataset manifest identifies each upstream source and license. CC-BY-4.0 content from reloading0101/threat-intelligence-dataset requires attribution. The GLM-4.7-Flash base model is distributed separately under its own license.