Model card: add Foundation-Sec-8B benchmark comparison (CyberMetric/CTI-Bench/SecQA, v4+v5)
Browse files
README.md
CHANGED
|
@@ -53,6 +53,47 @@ meaningful for this suite. Known v4 gap: `atk-sqli` — the model prefers to sta
|
|
| 53 |
staged vulnerable app and fuzz it over HTTP instead of reading the offline artifacts
|
| 54 |
(trace-verified behavioral prior, targeted in round 5).
|
| 55 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
## Usage
|
| 57 |
|
| 58 |
vLLM, short-task/eval serving (disable static YaRN):
|
|
|
|
| 53 |
staged vulnerable app and fuzz it over HTTP instead of reading the offline artifacts
|
| 54 |
(trace-verified behavioral prior, targeted in round 5).
|
| 55 |
|
| 56 |
+
## External benchmark: vs Cisco Foundation-Sec-8B (MCQ knowledge eval, 2026-08-01)
|
| 57 |
+
|
| 58 |
+
Independent comparison on 1,210 multiple-choice cybersecurity questions
|
| 59 |
+
(CyberMetric-500, CTI-Bench-MCQ 500/2500 sample, SecQA v1+v2 test; seed 42,
|
| 60 |
+
temperature 0). Each model was evaluated in its **native prompting mode**:
|
| 61 |
+
Foundation-Sec-8B is a *base* model and was evaluated with 3-shot raw completion
|
| 62 |
+
(its intended usage per Cisco's model card); k3-sec-8b was evaluated zero-shot
|
| 63 |
+
chat (Qwen3 template, non-thinking). Identical question sets for every model.
|
| 64 |
+
|
| 65 |
+
| Benchmark (n) | Foundation-Sec-8B | k3-sec-8b v4 | k3-sec-8b v5 |
|
| 66 |
+
|---|---:|---:|---:|
|
| 67 |
+
| CyberMetric-500 | 84.20% | **86.80%** | **86.80%** |
|
| 68 |
+
| CTI-Bench-MCQ (500) | **66.00%** | 59.20% | 59.80% |
|
| 69 |
+
| SecQA v1+v2 (210) | 98.10% | 98.10% | 98.10% |
|
| 70 |
+
| **Overall (1,210)** | **79.09%** | 77.36% | 77.60% |
|
| 71 |
+
|
| 72 |
+
Parse failures: 0 for all models in native modes.
|
| 73 |
+
|
| 74 |
+
**Question-level overlap, v5 vs Foundation-Sec** (exclusive wins): CyberMetric —
|
| 75 |
+
k3 +43 / F-Sec +30 · CTI-Bench — F-Sec +81 / k3 +50 · SecQA — 4 / 4 (both miss
|
| 76 |
+
the same 8 questions). **v4 → v5 flips:** +8 gained / −5 lost across all
|
| 77 |
+
benchmarks — knowledge MCQ is flat by design, since the v5 round targeted
|
| 78 |
+
agentic SQLi behavior, not knowledge.
|
| 79 |
+
|
| 80 |
+
**Prompting-mode sensitivity.** Foundation-Sec-8B collapses under zero-shot chat
|
| 81 |
+
prompting (~34% accuracy, ~40% whitespace/echo responses) and only performs via
|
| 82 |
+
few-shot completion — do not deploy it behind a naive chat template. k3-sec-8b
|
| 83 |
+
is robust across both modes (86.8% chat vs 87.4% completion on CyberMetric, both
|
| 84 |
+
versions) with 100% answer-format compliance.
|
| 85 |
+
|
| 86 |
+
**Qualitative** (6 open-ended security prompts): Foundation-Sec gives terse,
|
| 87 |
+
mostly-correct completion-style answers (~152 words avg); k3 v4/v5 give
|
| 88 |
+
structured markdown answers (~264/~307 words avg) with finer-grained precision
|
| 89 |
+
(e.g. LSASS dumping → T1003.001 sub-technique + detection methods, vs T1003 only).
|
| 90 |
+
|
| 91 |
+
**Takeaways.** k3-sec-8b (v4/v5) is stronger on broad general security knowledge
|
| 92 |
+
(CyberMetric) and far more usable interactively; Foundation-Sec-8B is stronger on
|
| 93 |
+
MITRE ATT&CK / threat-intel mapping (CTI-Bench) — the clearest data target for
|
| 94 |
+
the next SFT round. Serving stack: vLLM 0.23, bf16, `max_model_len=8192`.
|
| 95 |
+
Eval harness + raw per-question results: `sec-eval/` (drive_eval.py, results_*.json).
|
| 96 |
+
|
| 97 |
## Usage
|
| 98 |
|
| 99 |
vLLM, short-task/eval serving (disable static YaRN):
|