CyberSecQwen-4B、サイバーセキュリティベンチマークで8Bモデルに匹敵
Alibabaの研究者らがCyberSecQwen-4Bをリリース。この4Bパラメータモデルは、サイバータスクにおいてCiscoの8Bモデルの精度の97.3%を維持しつつ、コンシューマーGPUで動作可能。
Hugging Face released CyberSecQwen-4B, a 4-billion parameter model fine-tuned for defensive cybersecurity. The model achieves 97.3% of the CTI-RCM accuracy of Cisco's Foundation-Sec-Instruct-8B and exceeds its CTI-MCQ score by +8.7 points, despite using half the parameters. It runs efficiently on a single 12 GB consumer GPU or a single AMD Instinct MI300X 192 GB instance, enabling local processing of sensitive data like incident reports and malware samples without API calls. The training pipeline used full bf16 precision and FlashAttention-2 on the AMD MI300X.