Instruction tuning for threat intelligence analysis depends on high-quality data, but current methods for generating cybersecurity instruction data have no reliable way to assess quality, according to a new preprint on arXiv (2610.20284v1). The authors argue that this gap undermines the effectiveness of tuned models.

To address this, the paper introduces a quality-aware, optimal transport-guided data selection approach. The method appears to combine quality scoring with optimal transport techniques to choose training examples more efficiently, though the abstract is truncated and full details are not available in the announcement.

The work points toward more efficient instruction tuning for threat-intelligence tasks, where labeled data is scarce and noisy. If the proposed selection strategy holds up, it could reduce the cost of building capable analysis models.