Spaces:
Running
Running
File size: 28,516 Bytes
8e6cbe9 e4409fa 1cc9a2b e4409fa 1cc9a2b e4409fa 1cc9a2b e4409fa 1cc9a2b e4409fa 1cc9a2b e4409fa 8e6cbe9 e4409fa 8e6cbe9 e4409fa 8e6cbe9 e4409fa 16651e0 e4409fa 8e6cbe9 e4409fa 8e6cbe9 e4409fa 2f419c2 e4409fa 8e6cbe9 e4409fa 8e6cbe9 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 |
<!DOCTYPE html>
<html>
<head>
<title>Watermark Detector</title>
<meta name="viewport" content="width=device-width, initial-scale=1">
<link href="https://cdn.jsdelivr.net/npm/[email protected]/dist/css/bootstrap.min.css" rel="stylesheet">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/font/bootstrap-icons.css">
<link rel="stylesheet" href="{{ url_for('static', filename='styles.css') }}">
</head>
<body>
<div class="container">
<div class="d-flex justify-content-between align-items-center">
<h1>Interactive watermark detector</h1>
<div class="d-flex gap-2">
<button class="btn btn-outline-secondary" data-bs-toggle="modal" data-bs-target="#helpModal">
<i class="bi bi-question-circle"></i>
</button>
<button class="btn btn-outline-secondary" data-bs-toggle="modal" data-bs-target="#paramsModal">
<i class="bi bi-gear"></i>
</button>
</div>
</div>
<!-- Help Modal -->
<div class="modal fade" id="helpModal" tabindex="-1">
<div class="modal-dialog modal-lg">
<div class="modal-content">
<div class="modal-header">
<h5 class="modal-title">Watermark Detection Help</h5>
<button type="button" class="btn-close" data-bs-dismiss="modal"></button>
</div>
<div class="modal-body help-modal-body">
<h4>What is LLM watermarking?</h4>
<p>LLM watermarking is a technique that slightly modifies how language models generate text, making it possible to detect if text was generated by a specific AI model without visibly changing the text quality.</p>
<h4>How to use this demo</h4>
<ol>
<li>Enter a prompt in the top text area to generate watermarked text</li>
<li>The generated text will appear in the second text box</li>
<li>The text will be automatically analyzed to show which tokens (parts of text) were influenced by the watermark</li>
<li>The statistics at the bottom show the detection results</li>
<li>You can also paste any text in the second box to test if it contains a watermark</li>
</ol>
<h4>Detection Methods</h4>
<p><strong>Maryland</strong>: A token-level detection algorithm that analyzes how unexpected each token is, based on the paper "<a href="https://arxiv.org/abs/2301.10226" target="_blank">A Watermark for Large Language Models</a>" by Kirchenbauer et al.</p>
<p><strong>OpenAI</strong>: A similar watermarking method inspired by initial reports from OpenAI.</p>
<p><strong>Maryland Z-score</strong>: A worse variant of the Maryland detector that uses z-scores for statistical interpretation.</p>
<p><strong>OpenAI Z-score</strong>: A worse variant of the OpenAI detector that uses z-scores for statistical interpretation.</p>
<h4>Parameters Explained</h4>
<dl class="help-description-list">
<dt>Detector Type</dt>
<dd>The algorithm used to detect watermarks in the text. Different detectors perform better in different scenarios.</dd>
<dt>Seed</dt>
<dd>The random seed used for watermarking. The detector must use the same seed that was used when generating the text. In a real-world scenario, this would be kept private by the model provider.</dd>
<dt>N-gram Size</dt>
<dd>The number of previous tokens considered when choosing "greenlist" tokens. Larger values make the watermark less robust against edits but may improve text quality.</dd>
<dt>Delta</dt>
<dd>The bias added to "greenlist" tokens during generation. Higher values make the watermark stronger but might affect text quality. Typical values range from 1.0 to 5.0.</dd>
<dt>Temperature</dt>
<dd>Controls randomness in text generation. Higher values (e.g., 1.0) produce more diverse outputs; lower values (e.g., 0.2) make outputs more focused and deterministic.</dd>
</dl>
<h4>Understanding Results</h4>
<dl class="help-description-list">
<dt>Tokens</dt>
<dd>The total number of tokens in the analyzed text. Tokens are units of text that may represent words, parts of words, or punctuation.</dd>
<dt>Scored Tokens</dt>
<dd>The number of tokens that were actually evaluated by the detector (excludes first few tokens that don't have enough context).</dd>
<dt>Final Score</dt>
<dd>A measure of how likely the text contains a watermark. Higher scores indicate stronger evidence of watermarking.</dd>
<dt>P-value</dt>
<dd>The statistical significance of the detection. Lower values (especially p < 1e-6) indicate strong evidence that the text was watermarked. Values close to 0.5 suggest no watermark is present.</dd>
</dl>
<h4>Related Papers</h4>
<ul class="paper-references">
<li>
<a href="https://arxiv.org/abs/2301.10226" target="_blank">A Watermark for Large Language Models</a>
<span class="paper-authors">Kirchenbauer, et al. (2023)</span>
</li>
<li>
<a href="https://arxiv.org/abs/2308.00113" target="_blank">Three Bricks to Consolidate Watermarks for Large Language Models</a>
<span class="paper-authors">Fernandez, et al. (2023)</span>
</li>
</ul>
</div>
<div class="modal-footer">
<button type="button" class="btn btn-primary" data-bs-dismiss="modal">Close</button>
</div>
</div>
</div>
</div>
<!-- Advanced Parameters Modal -->
<div class="modal fade" id="paramsModal" tabindex="-1">
<div class="modal-dialog">
<div class="modal-content">
<div class="modal-header">
<h5 class="modal-title">Advanced Parameters</h5>
<button type="button" class="btn-close" data-bs-dismiss="modal"></button>
</div>
<div class="modal-body">
<div class="mb-3">
<div class="d-flex align-items-center">
<i class="bi bi-info-circle param-help-icon me-2" data-bs-toggle="tooltip" title="The algorithm used to detect watermarks in the text. Different algorithms have different strengths and weaknesses."></i>
<label for="detectorType" class="form-label mb-0">Detector Type</label>
</div>
<select class="form-select" id="detectorType">
<option value="maryland">Maryland</option>
<option value="marylandz">Maryland Z-score</option>
<option value="openai">OpenAI</option>
<option value="openaiz">OpenAI Z-score</option>
</select>
<div class="form-text">Type of watermark detection algorithm</div>
</div>
<div class="mb-3">
<div class="d-flex align-items-center">
<i class="bi bi-info-circle param-help-icon me-2" data-bs-toggle="tooltip" title="The random seed used for watermarking. For detection to work, this must match the seed used during generation."></i>
<label for="seed" class="form-label mb-0">Seed</label>
</div>
<input type="number" class="form-control" id="seed" value="0">
<div class="form-text">Random seed for the watermark detector</div>
</div>
<div class="mb-3">
<div class="d-flex align-items-center">
<i class="bi bi-info-circle param-help-icon me-2" data-bs-toggle="tooltip" title="The number of previous tokens considered when determining the hashing. Higher values are more secure but may affect robustness. Lower values may introduce a bias in the text distribution."></i>
<label for="ngram" class="form-label mb-0">N-gram Size</label>
</div>
<input type="number" class="form-control" id="ngram" value="1">
<div class="form-text">Size of the n-gram window used for detection</div>
</div>
<div class="mb-3">
<div class="d-flex align-items-center">
<i class="bi bi-info-circle param-help-icon me-2" data-bs-toggle="tooltip" title="The bias added to greenlist tokens during generation. Higher values (2.0-5.0) make watermarks easier to detect but might affect text quality."></i>
<label for="delta" class="form-label mb-0">Delta</label>
</div>
<input type="number" step="0.1" class="form-control" id="delta" value="2.0">
<div class="form-text">Bias added to greenlist tokens (for Maryland method)</div>
</div>
<div class="mb-3">
<div class="d-flex align-items-center">
<i class="bi bi-info-circle param-help-icon me-2" data-bs-toggle="tooltip" title="Controls randomness in generation. Higher values (closer to 2.0) create more diverse outputs and make watermarks easier to detect."></i>
<label for="temperature" class="form-label mb-0">Temperature</label>
</div>
<input type="number" step="0.1" class="form-control" id="temperature" value="0.8">
<div class="form-text">Temperature for sampling (higher = more random)</div>
</div>
</div>
<div class="modal-footer">
<button type="button" class="btn btn-secondary" data-bs-dismiss="modal">Close</button>
<button type="button" class="btn btn-primary" id="applyParams">Apply</button>
</div>
</div>
</div>
</div>
<!-- Input Form -->
<div class="input-section">
<div class="prompt-container">
<textarea id="prompt_text"
placeholder="Enter your prompt here to generate text with the model..."></textarea>
<button class="floating-btn" id="generateBtn">
<i class="bi bi-send-fill send-icon"></i>
<i class="bi bi-stop-fill stop-icon"></i>
</button>
</div>
<textarea id="user_text"
placeholder="Generated text will appear here. Replace or edit this text to see how watermark detection works."></textarea>
</div>
<!-- Token Display -->
<div class="token-display" id="tokenDisplay"></div>
<!-- Statistics -->
<div class="stats-container">
<div>
<div class="stat-value" id="tokenCount">0</div>
<div class="stat-label">
Tokens
<i class="bi bi-question-circle help-icon"></i>
<span class="help-tooltip">Total number of tokens in the text</span>
</div>
</div>
<div>
<div class="stat-value" id="scoredTokens">0</div>
<div class="stat-label">
Scored Tokens
<i class="bi bi-question-circle help-icon"></i>
<span class="help-tooltip">Number of tokens that were actually scored by the detector (excludes first n-gram tokens and duplicates)</span>
</div>
</div>
<div>
<div class="stat-value" id="finalScore">0.00</div>
<div class="stat-label">
Final Score
<i class="bi bi-question-circle help-icon"></i>
<span class="help-tooltip">Cumulative score from all scored tokens. Higher values indicate more likely watermarked text</span>
</div>
</div>
<div>
<div class="stat-value" id="pValue">0.500</div>
<div class="stat-label">
P-value
<i class="bi bi-question-circle help-icon"></i>
<span class="help-tooltip">Statistical significance of the score. Lower values indicate stronger evidence of watermarking (p < 0.05 is typically considered significant)</span>
</div>
</div>
</div>
</div>
<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/js/bootstrap.bundle.min.js"></script>
<script>
let debounceTimeout = null;
let abortController = null;
const textarea = document.getElementById('user_text');
const promptArea = document.getElementById('prompt_text');
const generateBtn = document.getElementById('generateBtn');
const tokenDisplay = document.getElementById('tokenDisplay');
const tokenCount = document.getElementById('tokenCount');
const scoredTokens = document.getElementById('scoredTokens');
const finalScore = document.getElementById('finalScore');
const pValue = document.getElementById('pValue');
const applyParamsBtn = document.getElementById('applyParams');
const seedInput = document.getElementById('seed');
const ngramInput = document.getElementById('ngram');
const detectorTypeSelect = document.getElementById('detectorType');
const deltaInput = document.getElementById('delta');
const temperatureInput = document.getElementById('temperature');
// Initialize tooltips
document.addEventListener('DOMContentLoaded', function() {
// Initialize Bootstrap tooltips
const tooltipTriggerList = [].slice.call(document.querySelectorAll('[data-bs-toggle="tooltip"]'));
const tooltipList = tooltipTriggerList.map(function (tooltipTriggerEl) {
return new bootstrap.Tooltip(tooltipTriggerEl);
});
// Initialize our tokenization
updateTokenization().catch(error => {
console.error('Error during initial tokenization:', error);
});
});
function startGeneration() {
const prompt = promptArea.value.trim();
if (!prompt) {
alert('Please enter a prompt first.');
return;
}
generateBtn.classList.add('generating');
textarea.value = '';
// Create new AbortController for this request
abortController = new AbortController();
// Get current parameters
const params = {
detector_type: detectorTypeSelect.value,
seed: parseInt(seedInput.value) || 0,
ngram: parseInt(ngramInput.value) || 1,
delta: parseFloat(deltaInput.value) || 2.0,
temperature: parseFloat(temperatureInput.value) || 0.8
};
// Create headers for SSE
const headers = new Headers({
'Content-Type': 'application/json',
'Accept': 'text/event-stream',
});
// Start fetch request with signal
fetch('/generate', {
method: 'POST',
headers: headers,
body: JSON.stringify({
prompt: prompt,
params: params
}),
signal: abortController.signal // Add the abort signal
}).then(response => {
const reader = response.body.getReader();
const decoder = new TextDecoder();
let buffer = '';
function processText(text) {
const lines = text.split('\n');
for (const line of lines) {
if (line.startsWith('data: ')) {
try {
const data = JSON.parse(line.slice(6));
if (data.error) {
alert('Error: ' + data.error);
stopGeneration();
return;
}
if (data.token) {
// Append new token to existing text
textarea.value += data.token;
updateTokenization();
}
if (data.text) {
// Final text (only used if something went wrong with streaming)
textarea.value = data.text;
updateTokenization();
}
if (data.done) {
stopGeneration();
}
} catch (e) {
console.error('Error parsing SSE data:', e);
}
}
}
}
function pump() {
return reader.read().then(({value, done}) => {
if (done) {
if (buffer.length > 0) {
processText(buffer);
}
return;
}
buffer += decoder.decode(value, {stream: true});
const lines = buffer.split('\n\n');
buffer = lines.pop();
for (const line of lines) {
processText(line);
}
return pump();
});
}
return pump();
})
.catch(error => {
if (error.name === 'AbortError') {
console.log('Generation stopped by user');
} else {
console.error('Error:', error);
alert('Error: Failed to generate text');
}
})
.finally(() => {
generateBtn.classList.remove('generating');
abortController = null;
});
}
function stopGeneration() {
if (abortController) {
abortController.abort();
abortController = null;
}
generateBtn.classList.remove('generating');
}
// Remove BOTH old event listeners and add just one new one
generateBtn.addEventListener('click', function(e) {
e.preventDefault(); // Prevent any double triggers
if (generateBtn.classList.contains('generating')) {
stopGeneration();
} else {
startGeneration();
}
});
async function updateTokenization() {
const text = textarea.value;
try {
// Validate parameters before sending
const seed = parseInt(seedInput.value);
const ngram = parseInt(ngramInput.value);
const delta = parseFloat(deltaInput.value);
const temperature = parseFloat(temperatureInput.value);
const response = await fetch('/tokenize', {
method: 'POST',
headers: {
'Content-Type': 'application/json',
},
body: JSON.stringify({
text: text,
params: {
detector_type: detectorTypeSelect.value,
seed: isNaN(seed) ? 0 : seed,
ngram: isNaN(ngram) ? 1 : ngram,
delta: isNaN(delta) ? 2.0 : delta,
temperature: isNaN(temperature) ? 0.8 : temperature
}
})
});
if (!response.ok) {
const errorData = await response.json();
throw new Error(errorData.error || `HTTP error! status: ${response.status}`);
}
const data = await response.json();
if (data.error) {
throw new Error(data.error);
}
// Update token display
tokenDisplay.innerHTML = data.tokens.map((token, i) => {
const score = data.scores[i];
const pvalue = data.pvalues[i];
const scoreDisplay = (score !== null && !isNaN(score)) ? score.toFixed(3) : 'N/A';
const pvalueDisplay = (pvalue !== null && !isNaN(pvalue)) ? formatPValue(pvalue) : 'N/A';
return `<span class="token" style="background-color: ${data.colors[i]}">
${token}
<div class="token-tooltip">
Score: ${scoreDisplay}<br>
P-value: ${pvalueDisplay}
</div>
</span>`;
}).join('');
// Update counts and stats - safely handle null values
tokenCount.textContent = data.token_count || 0;
scoredTokens.textContent = data.ntoks_scored || 0;
finalScore.textContent = (data.final_score !== null && !isNaN(data.final_score)) ?
data.final_score.toFixed(2) : '0.00';
pValue.textContent = (data.final_pvalue !== null && !isNaN(data.final_pvalue)) ?
formatPValue(data.final_pvalue) : '0.500';
// Clear any previous error
const existingError = tokenDisplay.querySelector('.alert-danger');
if (existingError) {
existingError.remove();
}
} catch (error) {
console.error('Error updating tokenization:', error);
// Show detailed error to user
tokenDisplay.innerHTML = `<div class="alert alert-danger">
<strong>Error:</strong> ${error.message || 'Error updating results. Please try again.'}
</div>`;
// Reset stats on error
tokenCount.textContent = '0';
scoredTokens.textContent = '0';
finalScore.textContent = '0.00';
pValue.textContent = '0.500';
}
}
// Increase debounce timeout and ensure it's properly cleared
textarea.addEventListener('input', function() {
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500); // Increased to 500ms
});
// Add input event listeners for parameter fields to trigger updates
seedInput.addEventListener('input', function() {
const value = this.value === '' ? '' : parseInt(this.value);
if (isNaN(value) && this.value !== '') {
this.value = "0";
}
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500);
});
ngramInput.addEventListener('input', function() {
const value = this.value === '' ? '' : parseInt(this.value);
if (isNaN(value) && this.value !== '') {
this.value = "1";
}
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500);
});
deltaInput.addEventListener('input', function() {
const value = this.value === '' ? '' : parseFloat(this.value);
if (isNaN(value) && this.value !== '') {
this.value = "2.0";
}
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500);
});
temperatureInput.addEventListener('input', function() {
const value = this.value === '' ? '' : parseFloat(this.value);
if (isNaN(value) && this.value !== '') {
this.value = "0.8";
}
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500);
});
// Add keyboard shortcut for applying changes
document.addEventListener('keydown', function(e) {
if ((e.metaKey || e.ctrlKey) && e.key === 'Enter') {
e.preventDefault();
if (document.activeElement === promptArea) {
if (generateBtn.classList.contains('generating')) {
stopGeneration();
} else {
startGeneration();
}
} else {
applyParamsBtn.click();
}
}
});
detectorTypeSelect.addEventListener('change', function() {
if (debounceTimeout) {
clearTimeout(debounceTimeout);
}
debounceTimeout = setTimeout(updateTokenization, 500);
});
// Ensure the modal apply button properly triggers an update
applyParamsBtn.addEventListener('click', function() {
updateTokenization().then(() => {
const modal = bootstrap.Modal.getInstance(document.getElementById('paramsModal'));
if (modal) {
modal.hide();
}
}).catch(error => {
console.error('Error applying parameters:', error);
});
});
// Add this helper function for formatting p-values
function formatPValue(value) {
if (value >= 0.001) {
return value.toFixed(3);
} else {
return value.toExponential(2);
}
}
</script>
</body>
</html>
|