Sensitivity Awareness on Access Denied Inc (ADI) benchmark (overall)
76.98AccuracyQwen3-8B (LoRA)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen3-8B (LoRA)Quantization=4-Bit, Fine-tuning protocol=LoRA, Model source=Open-Source2026.01 | 76.98 | 4.18 | 16.83 | 97.28 | 56.69 | 88.67 | 33.87 | |
| GPT-5 nanoModel source=Closed-Source, Precision=Full-precision2026.01 | 74.73 | 3.33 | 21.99 | 98.82 | 50.64 | 98 | 1.87 | |
| Qwen3-14B (LoRA)Quantization=4-Bit, Fine-tuning protocol=LoRA, Model source=Open-Source2026.01 | 73.72 | 0.68 | 21.17 | 97.33 | 50.11 | 84.53 | 15.33 | |
| Gemini 2.5 Flash LiteModel source=Closed-Source, Precision=Full-precision2026.01 | 63.3 | 0.4 | 33.91 | 97.22 | 29.39 | 84.93 | 0.8 | |
| Amazon Nova Lite-v1Model source=Closed-Source, Precision=Full-precision2026.01 | 63.3 | 5.89 | 30.91 | 94.08 | 30.65 | 92.93 | 0.4 | |
| Qwen3-14BQuantization=4-Bit, Fine-tuning protocol=Baseline, Model source=Open-Source2026.01 | 60.3 | 4.83 | 27.09 | 98.76 | 21.85 | 98.8 | 4.4 | |
| Mistral Nemo (12B)Model source=Open-Source, Precision=Full-precision, Parameters=12B2026.01 | 60.25 | 16.36 | 21.73 | 80.89 | 39.6 | 65.07 | 22.27 | |
| Llama 4 Scout (17B)Model source=Open-Source, Precision=Full-precision, Parameters=17B2026.01 | 58.19 | 1.28 | 40.39 | 99.41 | 16.97 | 96.27 | 6.67 | |
| Qwen3-8BQuantization=4-Bit, Fine-tuning protocol=Baseline, Model source=Open-Source2026.01 | 55.27 | 5.4 | 26.94 | 96 | 14.53 | 92.4 | 0.93 | |
| Phi-4 (14B)Model source=Open-Source, Precision=Full-precision, Parameters=14B2026.01 | 49.31 | 10.49 | 30.14 | 76.69 | 21.94 | 69.06 | 21.87 | |
| Llama 3.1 (8B)Model source=Open-Source, Precision=Full-precision, Parameters=8B2026.01 | 43.26 | 8.6 | 31.2 | 83.64 | 2.87 | 83.87 | 4.53 |