qwen2.5-7b-sigint-v1-GGUF
Version 1 of the mai-abliterator cybersecurity domain fine-tune.
Base: Qwen2.5-7B-Instruct | Method: QLoRA (rank 64) | Platform: Modal A100
Quantization: Q4_K_M (4.4GB) | Framework: Unsloth → llama.cpp GGUF
What this model does well
- CVE/kernel exploitation: +9.8% over baseline
- Embedded/IoT exploitation: +6.3%
- Drone/robotics security: +7.1%
- General intelligence retained (sanity check: code, knowledge, creative all ✓)
What this model does NOT do well
- Code generation: 0% — training data was 100% prose. The model learned to explain, not implement.
- SIGINT: -44.6% — RF physics, SDR workflows, signals intelligence training data too thin.
- Network attacks: -36.6% — WiFi/BT/LoRa attack surface coverage insufficient.
Benchmarked against baseline
| Metric | Baseline | v1 | Δ |
|---|
| Domain terms (avg) | 14.0 | 11.9 | -14.9% |
| Code present | 76% | 0% | -100% |
| CVE domain | 16.4 | 18.0 | +9.8% |
| SIGINT domain | 18.4 | 10.2 | -44.6% |
Full benchmark: 25 questions across CVE/SIGINT/DRONE/EMBED/NET domains.
Sanity check: reasoning, code generation, general knowledge — all preserved from base.
Use this model for
- Cybersecurity prose explanations
- CVE analysis
- Embedded system exploitation walkthroughs
- Understanding what NOT to do in v2 training data
Do NOT use this model for
- Writing exploit code → will refuse or produce prose
- SIGINT/RF analysis → hallucination risk
- Any production military contract work → v2 required
