Spaces:
Running
Running
Update src/App.js
Browse files- src/App.js +14 -9
src/App.js
CHANGED
|
@@ -95,11 +95,11 @@ const BenchmarkChart = () => {
|
|
| 95 |
const [currentPhase, setCurrentPhase] = useState('baseline');
|
| 96 |
const [currentMethodIndex, setCurrentMethodIndex] = useState(0);
|
| 97 |
|
| 98 |
-
const synthesisMethodsOrder = ['
|
| 99 |
|
| 100 |
const phases = [
|
| 101 |
{ key: 'baseline', label: 'Human Baseline ASR' },
|
| 102 |
-
{ key: 'additive_synthesis', label: 'Human +
|
| 103 |
];
|
| 104 |
|
| 105 |
useEffect(() => {
|
|
@@ -130,7 +130,7 @@ const BenchmarkChart = () => {
|
|
| 130 |
if (phase === 'baseline') {
|
| 131 |
return modelData.baseline;
|
| 132 |
} else if (phase === 'additive_synthesis') {
|
| 133 |
-
// Show the highest ASR achieved by any method tried so far
|
| 134 |
let maxASR = modelData.baseline;
|
| 135 |
|
| 136 |
for (let i = 0; i < currentMethodIndex; i++) {
|
|
@@ -158,7 +158,12 @@ const BenchmarkChart = () => {
|
|
| 158 |
}
|
| 159 |
|
| 160 |
if (availableMethods.length === 0) return 'Human Baseline';
|
| 161 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 162 |
};
|
| 163 |
|
| 164 |
const getBarColor = (modelData, phase) => {
|
|
@@ -177,7 +182,7 @@ const BenchmarkChart = () => {
|
|
| 177 |
{/* Header */}
|
| 178 |
<div className="text-center mb-6">
|
| 179 |
<h1 className="text-3xl font-bold text-white mb-3">
|
| 180 |
-
LLM Attack Success Rate with
|
| 181 |
</h1>
|
| 182 |
<p className="text-slate-300">
|
| 183 |
SafetyBench Aug 2025 - Real ASR Calculations
|
|
@@ -190,8 +195,8 @@ const BenchmarkChart = () => {
|
|
| 190 |
<div className="text-left">
|
| 191 |
<p className="text-yellow-200 font-semibold mb-2">Methodology Note</p>
|
| 192 |
<p className="text-yellow-100 text-sm leading-relaxed">
|
| 193 |
-
<strong>Additive Visualization:</strong> This chart shows cumulative impact by progressively adding each
|
| 194 |
-
Values >100% represent
|
| 195 |
Results are based on HarmBench Grading methodology and should be interpreted as relative performance indicators.
|
| 196 |
</p>
|
| 197 |
</div>
|
|
@@ -260,7 +265,7 @@ const BenchmarkChart = () => {
|
|
| 260 |
</div>
|
| 261 |
<div className="flex items-center space-x-2">
|
| 262 |
<div className="w-3 h-3 bg-gradient-to-r from-green-500 to-green-600 rounded"></div>
|
| 263 |
-
<span className="text-gray-700">With
|
| 264 |
</div>
|
| 265 |
</div>
|
| 266 |
</div>
|
|
@@ -271,7 +276,7 @@ const BenchmarkChart = () => {
|
|
| 271 |
Top performers: Grok 4 (85.15%), Deepseek R1-0528 (83.76%), Llama 3.1 405B (80.75%)
|
| 272 |
</p>
|
| 273 |
<p className="text-xs">
|
| 274 |
-
Shows highest ASR achieved when combining human attempts with
|
| 275 |
</p>
|
| 276 |
</div>
|
| 277 |
</div>
|
|
|
|
| 95 |
const [currentPhase, setCurrentPhase] = useState('baseline');
|
| 96 |
const [currentMethodIndex, setCurrentMethodIndex] = useState(0);
|
| 97 |
|
| 98 |
+
const synthesisMethodsOrder = ['keyword_objective_combined', 'root_problem'];
|
| 99 |
|
| 100 |
const phases = [
|
| 101 |
{ key: 'baseline', label: 'Human Baseline ASR' },
|
| 102 |
+
{ key: 'additive_synthesis', label: 'Human + Transformation Methods ASR' }
|
| 103 |
];
|
| 104 |
|
| 105 |
useEffect(() => {
|
|
|
|
| 130 |
if (phase === 'baseline') {
|
| 131 |
return modelData.baseline;
|
| 132 |
} else if (phase === 'additive_synthesis') {
|
| 133 |
+
// Show the highest ASR achieved by any transformation method tried so far
|
| 134 |
let maxASR = modelData.baseline;
|
| 135 |
|
| 136 |
for (let i = 0; i < currentMethodIndex; i++) {
|
|
|
|
| 158 |
}
|
| 159 |
|
| 160 |
if (availableMethods.length === 0) return 'Human Baseline';
|
| 161 |
+
|
| 162 |
+
const lastMethod = availableMethods[availableMethods.length - 1];
|
| 163 |
+
if (lastMethod === 'keyword_objective_combined') return 'Keyword/Objective Transformation';
|
| 164 |
+
if (lastMethod === 'root_problem') return 'Root Problem Transformation';
|
| 165 |
+
|
| 166 |
+
return lastMethod.replace(/_/g, ' ').replace(/\b\w/g, l => l.toUpperCase());
|
| 167 |
};
|
| 168 |
|
| 169 |
const getBarColor = (modelData, phase) => {
|
|
|
|
| 182 |
{/* Header */}
|
| 183 |
<div className="text-center mb-6">
|
| 184 |
<h1 className="text-3xl font-bold text-white mb-3">
|
| 185 |
+
LLM Attack Success Rate with Transformation Methods
|
| 186 |
</h1>
|
| 187 |
<p className="text-slate-300">
|
| 188 |
SafetyBench Aug 2025 - Real ASR Calculations
|
|
|
|
| 195 |
<div className="text-left">
|
| 196 |
<p className="text-yellow-200 font-semibold mb-2">Methodology Note</p>
|
| 197 |
<p className="text-yellow-100 text-sm leading-relaxed">
|
| 198 |
+
<strong>Additive Visualization:</strong> This chart shows cumulative impact by progressively adding each transformation method's individual attack success rate.
|
| 199 |
+
Values >100% represent transformation of multiple conversations off one failed, human seed conversation.
|
| 200 |
Results are based on HarmBench Grading methodology and should be interpreted as relative performance indicators.
|
| 201 |
</p>
|
| 202 |
</div>
|
|
|
|
| 265 |
</div>
|
| 266 |
<div className="flex items-center space-x-2">
|
| 267 |
<div className="w-3 h-3 bg-gradient-to-r from-green-500 to-green-600 rounded"></div>
|
| 268 |
+
<span className="text-gray-700">With Transformation Methods</span>
|
| 269 |
</div>
|
| 270 |
</div>
|
| 271 |
</div>
|
|
|
|
| 276 |
Top performers: Grok 4 (85.15%), Deepseek R1-0528 (83.76%), Llama 3.1 405B (80.75%)
|
| 277 |
</p>
|
| 278 |
<p className="text-xs">
|
| 279 |
+
Shows highest ASR achieved when combining human attempts with transformation methods
|
| 280 |
</p>
|
| 281 |
</div>
|
| 282 |
</div>
|