GYLH commited on
Commit
7e8150d
·
verified ·
1 Parent(s): bfaa95b

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +9 -6
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