Heimdall V1.1: 21,258 Cybersecurity Defense Instruction-Tuning Triples

21,258 high-quality system, user, and assistant triples for training alignment-safe, defensive cybersecurity large language models. The dataset was curated from over 100,000 technical sources, rigorously cleaned and filtered to enforce strict ethical boundaries, and is Apache-2.0 licensed. It was created by AlicanKiraz0 and last updated on June 21, 2025.

Use Cases

• Fine-tune large language models for defensive cybersecurity tasks based on the described instruction-response triples.
• Train models to generate safe and ethical responses to cybersecurity prompts based on the dataset's strict filtering.
• Benchmark LLM performance on cybersecurity frameworks like OWASP Top 10, NIST CSF, and MITRE ATT&CK based on the dataset's coverage.
• Develop AI assistants for cybersecurity education or threat analysis based on the curated technical Q&A format.

Strengths

• 21,258 instruction-response triples, representing a 760% increase from the previous version.
• Curated from over 100,000 technical sources and rigorously cleaned.
• Covers multiple cybersecurity frameworks including OWASP Top 10, NIST CSF, and MITRE ATT&CK.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but specific data format, file size, and sample data are unavailable for inspection.

Provenance

Source

huggingface user AlicanKiraz0

Collection Method

Curated from over 100,000 technical sources, then cleaned and filtered.

Time Range

Dataset version released June 21, 2025.

Freshness

Last updated 2025-06-21 16:04:10.

Geography

Region tag suggests 'us', but specific spatial coverage is not detailed.