Update README.md
Browse files
README.md
CHANGED
|
@@ -41,28 +41,31 @@ To this end, RADIUS-Drive introduces the SAR (Safety → Awareness → Reasoning
|
|
| 41 |
## 任务 / Tasks
|
| 42 |
### SAF(Safety)
|
| 43 |
|
| 44 |
-
检测模型是否能够识别并响应安全关键风险信号。
|
|
|
|
| 45 |
Detect whether the model correctly triggers safety awareness under risk-critical conditions.
|
| 46 |
|
| 47 |
### AWR(Awareness)
|
| 48 |
|
| 49 |
-
评估模型是否能够定位并解释风险来源,而非仅给出结果。
|
|
|
|
| 50 |
Assess whether the model localizes and explains the source of risk, beyond outcome correctness.
|
| 51 |
|
| 52 |
### REA(Reasoning)
|
| 53 |
|
| 54 |
-
验证模型的决策是否基于与风险一致的因果推理过程。
|
|
|
|
| 55 |
Verify whether decisions are supported by risk-consistent causal reasoning.
|
| 56 |
|
| 57 |
### X-CONS(Cross-Phase Consistency)
|
| 58 |
|
| 59 |
-
检查 Safety、Awareness 与 Reasoning 之间是否存在一致性违背(Pseudo-Correctness)。
|
|
|
|
| 60 |
Evaluate cross-phase consistency to diagnose pseudo-correct behavior.
|
| 61 |
|
| 62 |
## 关键约定 / Key Conventions
|
| 63 |
|
| 64 |
-
所有样本均配备 可审计的 ground truth,支持
|
| 65 |
-
reference-based injection 与 reference-free generation 两种场景构造方式。
|
| 66 |
|
| 67 |
All samples provide auditable ground truth, supporting both reference-based injection and reference-free generation.
|
| 68 |
|
|
|
|
| 41 |
## 任务 / Tasks
|
| 42 |
### SAF(Safety)
|
| 43 |
|
| 44 |
+
检测模型是否能够识别并响应安全关键风险信号。
|
| 45 |
+
|
| 46 |
Detect whether the model correctly triggers safety awareness under risk-critical conditions.
|
| 47 |
|
| 48 |
### AWR(Awareness)
|
| 49 |
|
| 50 |
+
评估模型是否能够定位并解释风险来源,而非仅给出结果。
|
| 51 |
+
|
| 52 |
Assess whether the model localizes and explains the source of risk, beyond outcome correctness.
|
| 53 |
|
| 54 |
### REA(Reasoning)
|
| 55 |
|
| 56 |
+
验证模型的决策是否基于与风险一致的因果推理过程。
|
| 57 |
+
|
| 58 |
Verify whether decisions are supported by risk-consistent causal reasoning.
|
| 59 |
|
| 60 |
### X-CONS(Cross-Phase Consistency)
|
| 61 |
|
| 62 |
+
检查 Safety、Awareness 与 Reasoning 之间是否存在一致性违背(Pseudo-Correctness)。
|
| 63 |
+
|
| 64 |
Evaluate cross-phase consistency to diagnose pseudo-correct behavior.
|
| 65 |
|
| 66 |
## 关键约定 / Key Conventions
|
| 67 |
|
| 68 |
+
所有样本均配备 可审计的 ground truth,支持reference-based injection 与 reference-free generation 两种场景构造方式。
|
|
|
|
| 69 |
|
| 70 |
All samples provide auditable ground truth, supporting both reference-based injection and reference-free generation.
|
| 71 |
|