Cybersecurity Threat Identification Instruction Dataset

32,000 instruction-following examples for training cybersecurity risk models. The dataset was curated by Vanessa Lopes and includes public reports and news, with outputs generated by GPT. It was last updated in April 2024.

Use Cases

• Fine-tune language models to generate threat classifications based on an 'instruction' and 'input' describing a security event.
• Train models to map cybersecurity incident descriptions from the 'input' field to standardized NIST taxonomy categories in the 'output'.
• Develop classifiers to assign IT risk classes from the EBA framework using the provided instruction-output pairs.
• Create synthetic training data pipelines by studying the relationship between the human-written 'instruction/input' and the GPT-generated 'output'.

Strengths

• 32,000 examples provide a substantial corpus for model training.
• Outputs are structured for specific cybersecurity taxonomies (NIST, EBA IT risk classes).

Limitations

• Dataset size is moderate for modern LLM fine-tuning; larger models may require additional data.
• Outputs are synthetic, generated by GPT, which may introduce biases or inaccuracies not present in human-labeled data.
• Limited metadata on the source, time range, and geographic origin of the underlying public reports and news.

Provenance

Source

Mix of public reports and news.

Collection Method

Curated by Vanessa Lopes; outputs generated by GPT.

Freshness

Last updated in April 2024.