| title: CyberSecQwen-4B · CTI Specialist | |
| emoji: "\U0001F6E1" | |
| colorFrom: blue | |
| colorTo: gray | |
| sdk: static | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Beating an 8B Cisco specialist at half the size | |
| hf_oauth: true | |
| hf_oauth_expiration_minutes: 480 | |
| # CyberSecQwen-4B | |
| A 4B-parameter Qwen3 fine-tune specialized for cyber threat intelligence — CWE classification, CVE-to-CWE mapping, code-pattern reasoning. Trained on a single AMD Instinct MI300X. | |
| | Benchmark | Score (n=5, temp 0.3) | vs. Foundation-Sec-Instruct-8B | | |
| |---|---|---| | |
| | CTI-RCM | 0.6664 ± 0.0023 | -1.9 pp at half the size | | |
| | CTI-MCQ | 0.5868 ± 0.0029 | **+8.7 pp at half the size** | | |
| This Space is the polished chat UI. It calls the ZeroGPU backend [`athena129/cybersecqwen-demo`](https://huggingface.co/spaces/athena129/cybersecqwen-demo) via `@gradio/client`. Cold start ~10–20 s, then warm. | |
| → Model card: [`athena129/CyberSecQwen-4B`](https://huggingface.co/athena129/CyberSecQwen-4B) · Source: [GitHub](https://github.com/GPT-64590/CyberSecQwen-4B) | |