Spaces:
Running
Running
File size: 12,129 Bytes
2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 5f3beb1 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 c1ae531 2607058 5f3beb1 2607058 5f3beb1 2607058 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 |
import React, { useState, useEffect } from 'react';
const BenchmarkChart = () => {
// Real data from your CSV
const benchmarkData = [
{
model: "Claude 4 Sonnet",
direct_conversation: 26.33,
keyword_objective_combined: 3.13
},
{
model: "Claude Opus 4.1",
direct_conversation: 20.67,
keyword_objective_combined: 3.65
},
{
model: "Deepseek R1-0528",
direct_conversation: 68.67,
keyword_objective_combined: 48.18
},
{
model: "GPT 5",
direct_conversation: 8.33,
keyword_objective_combined: 3.65,
bio_topic_change: 23.5,
enhancement: 10,
root_problem: 4.5
},
{
model: "GPT 5 mini",
direct_conversation: 7.67,
keyword_objective_combined: 3.91,
bio_topic_change: 14.5,
enhancement: 5.5,
root_problem: 3
},
{
model: "GPT o3",
direct_conversation: 22,
keyword_objective_combined: 10.94
},
{
model: "Gemini 2.5 Pro",
direct_conversation: 55.67,
keyword_objective_combined: 41.67,
bio_topic_change: 53.5,
enhancement: 47,
root_problem: 26
},
{
model: "Grok 4",
direct_conversation: 68.67,
keyword_objective_combined: 52.6
},
{
model: "Llama 3.1 405B",
direct_conversation: 67,
keyword_objective_combined: 41.67
}
];
const [currentPhase, setCurrentPhase] = useState('baseline');
const [currentMethodIndex, setCurrentMethodIndex] = useState(0);
const synthesisMethodsOrder = ['keyword_objective_combined', 'bio_topic_change', 'enhancement', 'root_problem'];
const phases = [
{ key: 'baseline', label: 'Direct Conversation (Baseline)' },
{ key: 'additive_synthesis', label: 'Adding Synthesis Methods' }
];
useEffect(() => {
const interval = setInterval(() => {
setCurrentPhase(prev => prev === 'baseline' ? 'additive_synthesis' : 'baseline');
setCurrentMethodIndex(0); // Reset when switching phases
}, 10000); // Reduced by 2 seconds: was 12000ms, now 10000ms
return () => clearInterval(interval);
}, []);
useEffect(() => {
if (currentPhase === 'additive_synthesis') {
const methodInterval = setInterval(() => {
setCurrentMethodIndex(prev => {
const nextIndex = prev + 1;
// Stay at final state (all methods added) for longer
if (nextIndex > synthesisMethodsOrder.length) {
return synthesisMethodsOrder.length; // Stay at max for longer
}
return nextIndex;
});
}, 2500); // Increased to 2.5 seconds to allow animation to complete
return () => clearInterval(methodInterval);
}
}, [currentPhase]);
const getCurrentValue = (modelData, phase) => {
if (phase === 'baseline') {
return modelData.direct_conversation || 0;
} else if (phase === 'additive_synthesis') {
let cumulativeValue = modelData.direct_conversation || 0;
// Add each synthesis method's contribution up to currentMethodIndex
for (let i = 0; i < currentMethodIndex; i++) {
const method = synthesisMethodsOrder[i];
if (modelData[method] !== undefined) {
cumulativeValue += modelData[method];
}
}
return cumulativeValue;
}
return 0;
};
const getCurrentMethodsAdded = (modelData, phase) => {
if (phase === 'baseline') return ['Direct Conversation'];
const methods = ['Direct Conversation'];
for (let i = 0; i < currentMethodIndex; i++) {
const method = synthesisMethodsOrder[i];
if (modelData[method] !== undefined) {
methods.push(method.replace(/_/g, ' ').replace(/\b\w/g, l => l.toUpperCase()));
}
}
return methods;
};
const getBarColor = (modelData, phase) => {
if (phase === 'baseline') {
return 'from-blue-500 to-blue-600';
} else if (phase === 'additive_synthesis' && currentMethodIndex > 0) {
// Only turn green when we've actually started adding methods
return 'from-green-500 to-green-600';
} else {
// Stay blue if we're in additive phase but haven't started adding yet
return 'from-blue-500 to-blue-600';
}
};
return (
<div className="min-h-screen bg-gradient-to-br from-slate-900 to-slate-800 p-8">
<div className="max-w-6xl mx-auto">
{/* Header */}
<div className="text-center mb-12">
<h1 className="text-4xl font-bold text-white mb-4">
LLM Safety Benchmark Results
</h1>
<p className="text-slate-300 text-lg">
SafetyBench Aug 2025 - Success Rate Comparison
</p>
{/* Methodology Disclaimer */}
<div className="mt-6 p-4 bg-yellow-900/30 border border-yellow-500/30 rounded-lg max-w-4xl mx-auto">
<div className="flex items-start space-x-3">
<div className="text-left">
<p className="text-yellow-200 font-semibold mb-2">Methodology Note</p>
<p className="text-yellow-100 text-sm leading-relaxed">
<strong>Additive Visualization:</strong> This chart shows cumulative impact by progressively adding each synthesis method's individual attack success rate.
Values >100% represent sythesis of multiple conversations off one failed, human seed conversation.
Results are based on HarmBench Grading methodology and should be interpreted as relative performance indicators.
</p>
</div>
</div>
</div>
<div className="mt-4 p-4 bg-slate-800 rounded-lg inline-block">
<p className="text-white font-semibold">
Current View: {phases.find(p => p.key === currentPhase)?.label}
</p>
{currentPhase === 'additive_synthesis' && currentMethodIndex > 0 && (
<p className="text-slate-300 text-sm mt-1">
Adding Method {currentMethodIndex}: {synthesisMethodsOrder[currentMethodIndex - 1]?.replace(/_/g, ' ').replace(/\b\w/g, l => l.toUpperCase())}
</p>
)}
</div>
</div>
{/* Chart Container */}
<div className="bg-white rounded-2xl shadow-2xl p-8">
<div className="space-y-8">
{benchmarkData.map((modelData, index) => {
const currentValue = getCurrentValue(modelData, currentPhase);
const baselineValue = modelData.direct_conversation;
const maxValue = 100; // Increased max scale since we're adding values
const barWidth = (currentValue / maxValue) * 100;
const methodsAdded = getCurrentMethodsAdded(modelData, currentPhase);
const totalGain = currentValue - baselineValue;
return (
<div key={modelData.model} className="relative">
{/* Model Name and Methods - Fixed Height Container */}
<div className="flex items-start justify-between mb-3 min-h-[60px]">
<div className="flex-1">
<h3 className="font-semibold text-gray-800 text-lg mb-1">
{modelData.model}
</h3>
<div className="text-sm text-gray-600 leading-tight">
<span className="font-medium">Methods: </span>
<span>{methodsAdded.join(' + ')}</span>
</div>
{/* Method breakdown - Always show, even during baseline */}
<div className="mt-1 min-h-[20px]">
{currentPhase === 'baseline' && (
<div className="text-xs text-gray-500 flex flex-wrap gap-x-4 gap-y-1">
<span className="text-blue-600">Base: {baselineValue.toFixed(1)}%</span>
<span className="text-gray-400">Ready to add synthesis methods...</span>
</div>
)}
{currentPhase === 'additive_synthesis' && (
<div className="text-xs text-gray-500 flex flex-wrap gap-x-4 gap-y-1">
<span className="text-blue-600">Base: {baselineValue.toFixed(1)}%</span>
{currentMethodIndex === 0 && (
<span className="text-gray-400">Starting to add methods...</span>
)}
{synthesisMethodsOrder.slice(0, currentMethodIndex).map(method => {
if (modelData[method] !== undefined) {
return (
<span key={method} className="text-green-600">
+{method.replace(/_/g, ' ')}: {modelData[method].toFixed(1)}%
</span>
);
}
return null;
})}
</div>
)}
</div>
</div>
<div className="text-right ml-4">
<span className="text-2xl font-bold text-gray-700">
{currentValue.toFixed(1)}%
</span>
{currentPhase === 'additive_synthesis' && totalGain > 0 && (
<div className="text-sm font-semibold text-green-600">
+{totalGain.toFixed(1)}% gain
</div>
)}
</div>
</div>
{/* Progress Bar */}
<div className="relative h-12 bg-gray-200 rounded-full overflow-hidden">
<div
className={`h-full bg-gradient-to-r ${getBarColor(modelData, currentPhase)} rounded-full transition-all duration-[2000ms] ease-out flex items-center justify-end pr-4`}
style={{ width: `${Math.max(barWidth, 5)}%` }}
>
<div className="text-white font-semibold text-sm">
{currentValue > 8 ? `${currentValue.toFixed(1)}%` : ''}
</div>
</div>
</div>
</div>
);
})}
</div>
{/* Legend */}
<div className="mt-8 flex justify-center space-x-6 flex-wrap">
<div className="flex items-center space-x-2">
<div className="w-4 h-4 bg-gradient-to-r from-blue-500 to-blue-600 rounded"></div>
<span className="text-gray-700">Baseline (Direct Conversation)</span>
</div>
<div className="flex items-center space-x-2">
<div className="w-4 h-4 bg-gradient-to-r from-green-500 to-green-600 rounded"></div>
<span className="text-gray-700">Adding Synthesis Methods</span>
</div>
</div>
</div>
{/* Footer Info */}
<div className="mt-8 text-center text-slate-400 space-y-2">
<p>Animation cycles every 10 seconds: baseline (4s) → progressive method addition (2.5s each) → smooth 2s bar growth</p>
<p className="text-sm">
Data from SafetyBench Aug 2025 • Synthesis methods test different attack vectors
</p>
<div className="text-xs mt-4 max-w-4xl mx-auto space-y-1">
<p><strong>Interpretation:</strong> Each synthesis method represents a different approach to testing model vulnerabilities</p>
<p><strong>Additive Display:</strong> Shows theoretical maximum impact when combining all available synthesis techniques</p>
<p><strong>Baseline:</strong> Direct conversation represents standard prompting without augmentation techniques</p>
</div>
</div>
</div>
</div>
);
};
export default BenchmarkChart; |