Commit ·
9c49fff
1
Parent(s): 33fed1a
Update extension UI, app pipeline and model loaders
Browse files- extension/popup.js +21 -6
- extension/sidepanel/sidepanel.js +20 -5
- src/app.py +44 -35
- src/js/editor.js +1 -1
- src/js/ui.js +8 -2
- src/model_loader.py +40 -188
- src/nlp/spelling/contextual_corrector.py +13 -19
- src/nlp/stage_locker.py +4 -0
extension/popup.js
CHANGED
|
@@ -55,6 +55,13 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 55 |
*/
|
| 56 |
let analyzedText = '';
|
| 57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
/**
|
| 59 |
* MED-2: Whether the analysis results are stale.
|
| 60 |
* Set to true when the user edits the textarea after analysis.
|
|
@@ -251,7 +258,7 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 251 |
btnApplyAll.classList.add('is-hidden');
|
| 252 |
if (analyzedText) {
|
| 253 |
suggestionsSection.classList.remove('is-hidden');
|
| 254 |
-
suggestionsList.innerHTML = '<div class="bayan-empty-state"><svg class="bayan-empty-state-icon" width="48" height="48" fill="none" stroke="currentColor" viewBox="0 0 24 24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 12l2 2 4-4m6 2a9 9 0 11-18 0 9 9 0 0118 0z"/></svg><p class="bayan-empty-state-
|
| 255 |
} else {
|
| 256 |
suggestionsSection.classList.add('is-hidden');
|
| 257 |
}
|
|
@@ -279,7 +286,9 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 279 |
if (!suggestion) return;
|
| 280 |
|
| 281 |
if (altText === suggestion.original) {
|
| 282 |
-
// Dismiss — remove from list,
|
|
|
|
|
|
|
| 283 |
currentSuggestions = removeSuggestion(currentSuggestions, suggestion.id);
|
| 284 |
} else {
|
| 285 |
// ═══════════════════════════════════════════════════
|
|
@@ -304,9 +313,13 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 304 |
});
|
| 305 |
});
|
| 306 |
|
| 307 |
-
// Show apply-all
|
| 308 |
-
|
| 309 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 310 |
}
|
| 311 |
|
| 312 |
// ══════════════════════════════════════════════════════════
|
|
@@ -349,7 +362,9 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 349 |
const data = await bayanAnalyze(text);
|
| 350 |
|
| 351 |
if (data.status === 'success' || data.status === 'partial') {
|
| 352 |
-
const suggestions = sortSuggestions(data.suggestions || [])
|
|
|
|
|
|
|
| 353 |
currentSuggestions = suggestions;
|
| 354 |
|
| 355 |
// MED-2: Snapshot the analyzed text — all offsets reference THIS string
|
|
|
|
| 55 |
*/
|
| 56 |
let analyzedText = '';
|
| 57 |
|
| 58 |
+
const _dismissedWords = new Set(
|
| 59 |
+
JSON.parse(localStorage.getItem('bayan_dismissed_words') || '[]')
|
| 60 |
+
);
|
| 61 |
+
function _saveDismissedWords() {
|
| 62 |
+
try { localStorage.setItem('bayan_dismissed_words', JSON.stringify([..._dismissedWords])); } catch {}
|
| 63 |
+
}
|
| 64 |
+
|
| 65 |
/**
|
| 66 |
* MED-2: Whether the analysis results are stale.
|
| 67 |
* Set to true when the user edits the textarea after analysis.
|
|
|
|
| 258 |
btnApplyAll.classList.add('is-hidden');
|
| 259 |
if (analyzedText) {
|
| 260 |
suggestionsSection.classList.remove('is-hidden');
|
| 261 |
+
suggestionsList.innerHTML = '<div class="bayan-empty-state"><svg class="bayan-empty-state-icon" width="48" height="48" fill="none" stroke="currentColor" viewBox="0 0 24 24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 12l2 2 4-4m6 2a9 9 0 11-18 0 9 9 0 0118 0z"/></svg><p class="bayan-empty-state-title">نصك ممتاز!</p><p class="bayan-empty-state-desc">لم نجد أي أخطاء — أحسنت! ✨</p></div>';
|
| 262 |
} else {
|
| 263 |
suggestionsSection.classList.add('is-hidden');
|
| 264 |
}
|
|
|
|
| 286 |
if (!suggestion) return;
|
| 287 |
|
| 288 |
if (altText === suggestion.original) {
|
| 289 |
+
// Dismiss — remove from list, add to permanent whitelist
|
| 290 |
+
_dismissedWords.add(suggestion.original);
|
| 291 |
+
_saveDismissedWords();
|
| 292 |
currentSuggestions = removeSuggestion(currentSuggestions, suggestion.id);
|
| 293 |
} else {
|
| 294 |
// ═══════════════════════════════════════════════════
|
|
|
|
| 313 |
});
|
| 314 |
});
|
| 315 |
|
| 316 |
+
// Show apply-all only when more than 1 suggestion
|
| 317 |
+
if (suggestions.length > 1) {
|
| 318 |
+
btnApplyAll.textContent = 'تطبيق الكل (' + suggestions.length.toLocaleString('ar-EG') + ')';
|
| 319 |
+
btnApplyAll.classList.remove('is-hidden');
|
| 320 |
+
} else {
|
| 321 |
+
btnApplyAll.classList.add('is-hidden');
|
| 322 |
+
}
|
| 323 |
}
|
| 324 |
|
| 325 |
// ══════════════════════════════════════════════════════════
|
|
|
|
| 362 |
const data = await bayanAnalyze(text);
|
| 363 |
|
| 364 |
if (data.status === 'success' || data.status === 'partial') {
|
| 365 |
+
const suggestions = sortSuggestions(data.suggestions || []).filter(
|
| 366 |
+
s => !_dismissedWords.has(s.original)
|
| 367 |
+
);
|
| 368 |
currentSuggestions = suggestions;
|
| 369 |
|
| 370 |
// MED-2: Snapshot the analyzed text — all offsets reference THIS string
|
extension/sidepanel/sidepanel.js
CHANGED
|
@@ -62,6 +62,13 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 62 |
let sourceSelectionText = '';
|
| 63 |
let summaryMode = 'paragraph';
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
const SCORE_CIRCUMFERENCE = 440;
|
| 66 |
|
| 67 |
// ══════════════════════════════════════════════════════════
|
|
@@ -206,7 +213,7 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 206 |
btnApplyAll.classList.add('is-hidden');
|
| 207 |
if (analyzedText) {
|
| 208 |
suggestionsSection.classList.remove('is-hidden');
|
| 209 |
-
suggestionsList.innerHTML = '<div class="sp-empty-state"><svg class="sp-empty-state-icon" width="48" height="48" fill="none" stroke="currentColor" viewBox="0 0 24 24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 12l2 2 4-4m6 2a9 9 0 11-18 0 9 9 0 0118 0z"/></svg><p class="sp-empty-state-
|
| 210 |
} else {
|
| 211 |
suggestionsSection.classList.add('is-hidden');
|
| 212 |
}
|
|
@@ -231,6 +238,8 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 231 |
if (!suggestion) return;
|
| 232 |
|
| 233 |
if (altText === suggestion.original) {
|
|
|
|
|
|
|
| 234 |
currentSuggestions = removeSuggestion(currentSuggestions, suggestion.id);
|
| 235 |
} else {
|
| 236 |
const result = applyAndRebase(analyzedText, suggestion, altText, currentSuggestions);
|
|
@@ -248,9 +257,13 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 248 |
});
|
| 249 |
});
|
| 250 |
|
| 251 |
-
// Show apply-all
|
| 252 |
-
|
| 253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 254 |
}
|
| 255 |
|
| 256 |
// ══════════════════════════════════════════════════════════
|
|
@@ -274,7 +287,9 @@ document.addEventListener('DOMContentLoaded', () => {
|
|
| 274 |
const data = await bayanAnalyze(text);
|
| 275 |
|
| 276 |
if (data.status === 'success' || data.status === 'partial') {
|
| 277 |
-
const suggestions = sortSuggestions(data.suggestions || [])
|
|
|
|
|
|
|
| 278 |
currentSuggestions = suggestions;
|
| 279 |
analyzedText = data.original;
|
| 280 |
|
|
|
|
| 62 |
let sourceSelectionText = '';
|
| 63 |
let summaryMode = 'paragraph';
|
| 64 |
|
| 65 |
+
const _dismissedWords = new Set(
|
| 66 |
+
JSON.parse(localStorage.getItem('bayan_dismissed_words') || '[]')
|
| 67 |
+
);
|
| 68 |
+
function _saveDismissedWords() {
|
| 69 |
+
try { localStorage.setItem('bayan_dismissed_words', JSON.stringify([..._dismissedWords])); } catch {}
|
| 70 |
+
}
|
| 71 |
+
|
| 72 |
const SCORE_CIRCUMFERENCE = 440;
|
| 73 |
|
| 74 |
// ══════════════════════════════════════════════════════════
|
|
|
|
| 213 |
btnApplyAll.classList.add('is-hidden');
|
| 214 |
if (analyzedText) {
|
| 215 |
suggestionsSection.classList.remove('is-hidden');
|
| 216 |
+
suggestionsList.innerHTML = '<div class="sp-empty-state"><svg class="sp-empty-state-icon" width="48" height="48" fill="none" stroke="currentColor" viewBox="0 0 24 24"><path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M9 12l2 2 4-4m6 2a9 9 0 11-18 0 9 9 0 0118 0z"/></svg><p class="sp-empty-state-title">نصك ممتاز!</p><p class="sp-empty-state-desc">لم نجد أي أخطاء — أحسنت! ✨</p></div>';
|
| 217 |
} else {
|
| 218 |
suggestionsSection.classList.add('is-hidden');
|
| 219 |
}
|
|
|
|
| 238 |
if (!suggestion) return;
|
| 239 |
|
| 240 |
if (altText === suggestion.original) {
|
| 241 |
+
_dismissedWords.add(suggestion.original);
|
| 242 |
+
_saveDismissedWords();
|
| 243 |
currentSuggestions = removeSuggestion(currentSuggestions, suggestion.id);
|
| 244 |
} else {
|
| 245 |
const result = applyAndRebase(analyzedText, suggestion, altText, currentSuggestions);
|
|
|
|
| 257 |
});
|
| 258 |
});
|
| 259 |
|
| 260 |
+
// Show apply-all only when more than 1 suggestion
|
| 261 |
+
if (suggestions.length > 1) {
|
| 262 |
+
btnApplyAll.textContent = 'تطبيق الكل (' + suggestions.length.toLocaleString('ar-EG') + ')';
|
| 263 |
+
btnApplyAll.classList.remove('is-hidden');
|
| 264 |
+
} else {
|
| 265 |
+
btnApplyAll.classList.add('is-hidden');
|
| 266 |
+
}
|
| 267 |
}
|
| 268 |
|
| 269 |
// ══════════════════════════════════════════════════════════
|
|
|
|
| 287 |
const data = await bayanAnalyze(text);
|
| 288 |
|
| 289 |
if (data.status === 'success' || data.status === 'partial') {
|
| 290 |
+
const suggestions = sortSuggestions(data.suggestions || []).filter(
|
| 291 |
+
s => !_dismissedWords.has(s.original)
|
| 292 |
+
);
|
| 293 |
currentSuggestions = suggestions;
|
| 294 |
analyzedText = data.original;
|
| 295 |
|
src/app.py
CHANGED
|
@@ -46,12 +46,10 @@ from model_loader import (
|
|
| 46 |
SummarizationModel,
|
| 47 |
SpellingModel,
|
| 48 |
AutocompleteModel,
|
| 49 |
-
GrammarModel,
|
| 50 |
PunctuationModel,
|
| 51 |
SUMMARIZATION_PATH,
|
| 52 |
SPELLING_PATH,
|
| 53 |
AUTOCOMPLETE_PATH,
|
| 54 |
-
GRAMMAR_PATH,
|
| 55 |
PUNCTUATION_PATH
|
| 56 |
)
|
| 57 |
|
|
@@ -754,24 +752,34 @@ class OffsetMapper:
|
|
| 754 |
inverted range (start > end) due to non-monotonic edits,
|
| 755 |
the end is clamped to max(new_start, new_end).
|
| 756 |
"""
|
| 757 |
-
new_start = self._forward_map_pos(start_in_before)
|
| 758 |
-
new_end = self._forward_map_pos(end_in_before)
|
| 759 |
# Monotonicity guard: prevent inverted ranges
|
| 760 |
new_end = max(new_start, new_end)
|
| 761 |
return new_start, new_end
|
| 762 |
|
| 763 |
-
def _forward_map_pos(self, pos):
|
| 764 |
"""Map a single position text_before → text_after. PRIVATE."""
|
|
|
|
| 765 |
for i1, i2, j1, j2 in self._opcodes:
|
| 766 |
if i1 <= pos <= i2:
|
| 767 |
-
|
| 768 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 769 |
ratio = (pos - i1) / (i2 - i1)
|
| 770 |
-
|
| 771 |
-
|
| 772 |
-
|
| 773 |
-
return last[3] + (pos - last[1])
|
| 774 |
-
return pos
|
| 775 |
|
| 776 |
|
| 777 |
|
|
@@ -1521,7 +1529,7 @@ def quran_verify():
|
|
| 1521 |
if not logger_quran_ok:
|
| 1522 |
return jsonify({'error': 'Quran search module not available'}), 503
|
| 1523 |
|
| 1524 |
-
data = request.get_json(
|
| 1525 |
text = data.get('text', '').strip()
|
| 1526 |
language = data.get('language', 'تدقيق الايات').strip()
|
| 1527 |
|
|
@@ -2226,8 +2234,10 @@ def analyze_text():
|
|
| 2226 |
# Replace structured content with Arabic placeholder tokens
|
| 2227 |
if _structured_placeholders:
|
| 2228 |
_structured_placeholders.sort(key=lambda x: x[0], reverse=True)
|
|
|
|
|
|
|
| 2229 |
for _sp_idx, (_sp_start, _sp_end, _sp_text) in enumerate(_structured_placeholders):
|
| 2230 |
-
_grammar_input_text = _grammar_input_text[:_sp_start] + f'
|
| 2231 |
logger.info(f"[ANALYZE] Protected {len(_structured_placeholders)} structured elements")
|
| 2232 |
|
| 2233 |
# 2. Grammar (runs on spelling-corrected text — word-level dependency)
|
|
@@ -2249,7 +2259,7 @@ def analyze_text():
|
|
| 2249 |
# FIX-03: Restore structured content in grammar output
|
| 2250 |
if _structured_placeholders:
|
| 2251 |
for _sp_idx, (_sp_start, _sp_end, _sp_text) in enumerate(_structured_placeholders):
|
| 2252 |
-
corrected_grammar = corrected_grammar.replace(f'
|
| 2253 |
|
| 2254 |
if corrected_grammar != ctx.current_text:
|
| 2255 |
diffs = get_word_diffs(ctx.current_text, corrected_grammar)
|
|
@@ -2504,19 +2514,6 @@ def analyze_text():
|
|
| 2504 |
)
|
| 2505 |
continue
|
| 2506 |
|
| 2507 |
-
# ── FIX-27a: Grammar structured data protection ──
|
| 2508 |
-
# Block grammar diffs where the original contains digits.
|
| 2509 |
-
# The grammar model corrupts dates/numbers/times/percentages.
|
| 2510 |
-
# e.g., '2026-06-22' → 'عشرين 26-06-22ا'
|
| 2511 |
-
if orig_text and any(c.isdigit() for c in orig_text):
|
| 2512 |
-
logger.info(
|
| 2513 |
-
f"[GRAMMAR] Blocked digit-containing diff: "
|
| 2514 |
-
f"'{orig_text}'\u2192'{corr_text}'"
|
| 2515 |
-
)
|
| 2516 |
-
logger.info(f'[FILTER-TEL] {_tel_json.dumps({"event":"filter_reject","filter":"DigitGuard","original":orig_text[:80],"correction":corr_text[:80]})}')
|
| 2517 |
-
_tel_events.append({"event":"filter_reject","filter":"DigitGuard","original":orig_text[:80],"correction":corr_text[:80]})
|
| 2518 |
-
continue
|
| 2519 |
-
|
| 2520 |
# ── FIX-27b: Grammar hallucination guard (Jaccard) ──
|
| 2521 |
# Block grammar diffs where the correction is too different
|
| 2522 |
# from the original (character-level Jaccard < 0.5).
|
|
@@ -2575,10 +2572,13 @@ def analyze_text():
|
|
| 2575 |
# FIX-22: Protect tanween (preserve ً ٌ ٍ from original)
|
| 2576 |
_TANWEEN_CHARS = set('ًٌٍ')
|
| 2577 |
if any(c in _TANWEEN_CHARS for c in orig_text) and not any(c in _TANWEEN_CHARS for c in corr_text):
|
| 2578 |
-
# Grammar stripped tanween — reattach
|
| 2579 |
for _tc in _TANWEEN_CHARS:
|
| 2580 |
if _tc in orig_text and _tc not in corr_text:
|
| 2581 |
-
|
|
|
|
|
|
|
|
|
|
| 2582 |
d['correction'] = corr_text
|
| 2583 |
break
|
| 2584 |
|
|
@@ -2609,14 +2609,25 @@ def analyze_text():
|
|
| 2609 |
logger.info(
|
| 2610 |
f"[GRAMMAR] Bracket-unbalanced output detected: "
|
| 2611 |
f"orig=({orig_opens},{orig_closes}), corr=({corr_opens},{corr_closes}). "
|
| 2612 |
-
f"
|
| 2613 |
)
|
|
|
|
| 2614 |
if _grammar_accepted_diffs:
|
| 2615 |
# FIX-04: Rebuild grammar text from ACCEPTED diffs only,
|
| 2616 |
# not the full model output. Prevents phantom corrections.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2617 |
_safe_grammar = ctx.current_text
|
| 2618 |
-
|
| 2619 |
-
for _ad in sorted(_grammar_accepted_diffs, key=lambda x: x['start'], reverse=True):
|
| 2620 |
_safe_grammar = (_safe_grammar[:_ad['start']] +
|
| 2621 |
_ad['correction'] +
|
| 2622 |
_safe_grammar[_ad['end']:])
|
|
@@ -2630,8 +2641,6 @@ def analyze_text():
|
|
| 2630 |
|
| 2631 |
# 3. Punctuation (runs on grammar-corrected text — PuncAra-v1 local model)
|
| 2632 |
# FIX-07: Skip punctuation for religious text
|
| 2633 |
-
# FIX-51: Skip punctuation when spelling+grammar found no errors (clean text)
|
| 2634 |
-
_has_real_corrections = any(p.stage in ('spelling', 'grammar') for p in ctx.patches.patches)
|
| 2635 |
if not _skip_all_stages:
|
| 2636 |
try:
|
| 2637 |
t0 = time.time()
|
|
|
|
| 46 |
SummarizationModel,
|
| 47 |
SpellingModel,
|
| 48 |
AutocompleteModel,
|
|
|
|
| 49 |
PunctuationModel,
|
| 50 |
SUMMARIZATION_PATH,
|
| 51 |
SPELLING_PATH,
|
| 52 |
AUTOCOMPLETE_PATH,
|
|
|
|
| 53 |
PUNCTUATION_PATH
|
| 54 |
)
|
| 55 |
|
|
|
|
| 752 |
inverted range (start > end) due to non-monotonic edits,
|
| 753 |
the end is clamped to max(new_start, new_end).
|
| 754 |
"""
|
| 755 |
+
new_start = self._forward_map_pos(start_in_before, is_end=False)
|
| 756 |
+
new_end = self._forward_map_pos(end_in_before, is_end=True)
|
| 757 |
# Monotonicity guard: prevent inverted ranges
|
| 758 |
new_end = max(new_start, new_end)
|
| 759 |
return new_start, new_end
|
| 760 |
|
| 761 |
+
def _forward_map_pos(self, pos, is_end=False):
|
| 762 |
"""Map a single position text_before → text_after. PRIVATE."""
|
| 763 |
+
matches = []
|
| 764 |
for i1, i2, j1, j2 in self._opcodes:
|
| 765 |
if i1 <= pos <= i2:
|
| 766 |
+
matches.append((i1, i2, j1, j2))
|
| 767 |
+
|
| 768 |
+
if not matches:
|
| 769 |
+
if self._opcodes:
|
| 770 |
+
last = self._opcodes[-1]
|
| 771 |
+
return last[3] + (pos - last[1])
|
| 772 |
+
return pos
|
| 773 |
+
|
| 774 |
+
mapped = []
|
| 775 |
+
for i1, i2, j1, j2 in matches:
|
| 776 |
+
if i2 == i1: # insertion in text_after
|
| 777 |
+
mapped.append(j2 if is_end else j1)
|
| 778 |
+
else:
|
| 779 |
ratio = (pos - i1) / (i2 - i1)
|
| 780 |
+
mapped.append(round(j1 + ratio * (j2 - j1)))
|
| 781 |
+
|
| 782 |
+
return max(mapped) if is_end else min(mapped)
|
|
|
|
|
|
|
| 783 |
|
| 784 |
|
| 785 |
|
|
|
|
| 1529 |
if not logger_quran_ok:
|
| 1530 |
return jsonify({'error': 'Quran search module not available'}), 503
|
| 1531 |
|
| 1532 |
+
data = request.get_json()
|
| 1533 |
text = data.get('text', '').strip()
|
| 1534 |
language = data.get('language', 'تدقيق الايات').strip()
|
| 1535 |
|
|
|
|
| 2234 |
# Replace structured content with Arabic placeholder tokens
|
| 2235 |
if _structured_placeholders:
|
| 2236 |
_structured_placeholders.sort(key=lambda x: x[0], reverse=True)
|
| 2237 |
+
import uuid as _uuid_struct
|
| 2238 |
+
_ph_prefix = _uuid_struct.uuid4().hex[:8]
|
| 2239 |
for _sp_idx, (_sp_start, _sp_end, _sp_text) in enumerate(_structured_placeholders):
|
| 2240 |
+
_grammar_input_text = _grammar_input_text[:_sp_start] + f'ـ{_ph_prefix}{_sp_idx}ـ' + _grammar_input_text[_sp_end:]
|
| 2241 |
logger.info(f"[ANALYZE] Protected {len(_structured_placeholders)} structured elements")
|
| 2242 |
|
| 2243 |
# 2. Grammar (runs on spelling-corrected text — word-level dependency)
|
|
|
|
| 2259 |
# FIX-03: Restore structured content in grammar output
|
| 2260 |
if _structured_placeholders:
|
| 2261 |
for _sp_idx, (_sp_start, _sp_end, _sp_text) in enumerate(_structured_placeholders):
|
| 2262 |
+
corrected_grammar = corrected_grammar.replace(f'ـ{_ph_prefix}{_sp_idx}ـ', _sp_text)
|
| 2263 |
|
| 2264 |
if corrected_grammar != ctx.current_text:
|
| 2265 |
diffs = get_word_diffs(ctx.current_text, corrected_grammar)
|
|
|
|
| 2514 |
)
|
| 2515 |
continue
|
| 2516 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2517 |
# ── FIX-27b: Grammar hallucination guard (Jaccard) ──
|
| 2518 |
# Block grammar diffs where the correction is too different
|
| 2519 |
# from the original (character-level Jaccard < 0.5).
|
|
|
|
| 2572 |
# FIX-22: Protect tanween (preserve ً ٌ ٍ from original)
|
| 2573 |
_TANWEEN_CHARS = set('ًٌٍ')
|
| 2574 |
if any(c in _TANWEEN_CHARS for c in orig_text) and not any(c in _TANWEEN_CHARS for c in corr_text):
|
| 2575 |
+
# Grammar stripped tanween — reattach before trailing punctuation
|
| 2576 |
for _tc in _TANWEEN_CHARS:
|
| 2577 |
if _tc in orig_text and _tc not in corr_text:
|
| 2578 |
+
_tn_punct_idx = len(corr_text)
|
| 2579 |
+
while _tn_punct_idx > 0 and corr_text[_tn_punct_idx - 1] in '.،؛؟!?!':
|
| 2580 |
+
_tn_punct_idx -= 1
|
| 2581 |
+
corr_text = corr_text[:_tn_punct_idx] + _tc + corr_text[_tn_punct_idx:]
|
| 2582 |
d['correction'] = corr_text
|
| 2583 |
break
|
| 2584 |
|
|
|
|
| 2609 |
logger.info(
|
| 2610 |
f"[GRAMMAR] Bracket-unbalanced output detected: "
|
| 2611 |
f"orig=({orig_opens},{orig_closes}), corr=({corr_opens},{corr_closes}). "
|
| 2612 |
+
f"Rejecting all grammar diffs."
|
| 2613 |
)
|
| 2614 |
+
_grammar_accepted_diffs = []
|
| 2615 |
if _grammar_accepted_diffs:
|
| 2616 |
# FIX-04: Rebuild grammar text from ACCEPTED diffs only,
|
| 2617 |
# not the full model output. Prevents phantom corrections.
|
| 2618 |
+
# Remove intra-stage overlaps: keep higher-start diff on conflict
|
| 2619 |
+
_grammar_accepted_diffs.sort(key=lambda x: x['start'])
|
| 2620 |
+
_non_overlapping = []
|
| 2621 |
+
for _ad in _grammar_accepted_diffs:
|
| 2622 |
+
if _non_overlapping and _ad['start'] < _non_overlapping[-1]['end']:
|
| 2623 |
+
logger.info(
|
| 2624 |
+
f"[GRAMMAR] Dropped overlapping diff [{_ad['start']}:{_ad['end']}] "
|
| 2625 |
+
f"(overlaps [{_non_overlapping[-1]['start']}:{_non_overlapping[-1]['end']}])"
|
| 2626 |
+
)
|
| 2627 |
+
continue
|
| 2628 |
+
_non_overlapping.append(_ad)
|
| 2629 |
_safe_grammar = ctx.current_text
|
| 2630 |
+
for _ad in reversed(_non_overlapping):
|
|
|
|
| 2631 |
_safe_grammar = (_safe_grammar[:_ad['start']] +
|
| 2632 |
_ad['correction'] +
|
| 2633 |
_safe_grammar[_ad['end']:])
|
|
|
|
| 2641 |
|
| 2642 |
# 3. Punctuation (runs on grammar-corrected text — PuncAra-v1 local model)
|
| 2643 |
# FIX-07: Skip punctuation for religious text
|
|
|
|
|
|
|
| 2644 |
if not _skip_all_stages:
|
| 2645 |
try:
|
| 2646 |
t0 = time.time()
|
src/js/editor.js
CHANGED
|
@@ -300,7 +300,7 @@ async function analyzeText() {
|
|
| 300 |
|
| 301 |
const data = await response.json();
|
| 302 |
|
| 303 |
-
if (data.status !== 'success' || !data.suggestions) {
|
| 304 |
renderWithoutSuggestions(text);
|
| 305 |
return;
|
| 306 |
}
|
|
|
|
| 300 |
|
| 301 |
const data = await response.json();
|
| 302 |
|
| 303 |
+
if ((data.status !== 'success' && data.status !== 'partial') || !data.suggestions) {
|
| 304 |
renderWithoutSuggestions(text);
|
| 305 |
return;
|
| 306 |
}
|
src/js/ui.js
CHANGED
|
@@ -26,7 +26,9 @@ function updateWritingScore(spelling, grammar, punctuation) {
|
|
| 26 |
const hintEl = document.getElementById('score-hint');
|
| 27 |
|
| 28 |
if (valueEl) {
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
? score.toLocaleString('ar-EG')
|
| 31 |
: '--';
|
| 32 |
}
|
|
@@ -38,7 +40,11 @@ function updateWritingScore(spelling, grammar, punctuation) {
|
|
| 38 |
|
| 39 |
if (hintEl) {
|
| 40 |
const total = spelling + grammar + punctuation;
|
| 41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
hintEl.innerHTML = 'ابدأ الكتابة لرؤية تقييمك<br><span class="text-xs">تحسين القواعد يرفع التقييم</span>';
|
| 43 |
} else if (score >= 90) {
|
| 44 |
hintEl.textContent = 'كتابة ممتازة! استمر.';
|
|
|
|
| 26 |
const hintEl = document.getElementById('score-hint');
|
| 27 |
|
| 28 |
if (valueEl) {
|
| 29 |
+
const editorText = typeof getEditorText === 'function' ? getEditorText() : '';
|
| 30 |
+
const hasText = editorText.trim().length > 0;
|
| 31 |
+
valueEl.textContent = (hasText || (spelling + grammar + punctuation) > 0)
|
| 32 |
? score.toLocaleString('ar-EG')
|
| 33 |
: '--';
|
| 34 |
}
|
|
|
|
| 40 |
|
| 41 |
if (hintEl) {
|
| 42 |
const total = spelling + grammar + punctuation;
|
| 43 |
+
const editorText = typeof getEditorText === 'function' ? getEditorText() : '';
|
| 44 |
+
const hasText = editorText.trim().length > 0;
|
| 45 |
+
if (total === 0 && hasText) {
|
| 46 |
+
hintEl.textContent = 'كتابة ممتازة! استمر.';
|
| 47 |
+
} else if (total === 0) {
|
| 48 |
hintEl.innerHTML = 'ابدأ الكتابة لرؤية تقييمك<br><span class="text-xs">تحسين القواعد يرفع التقييم</span>';
|
| 49 |
} else if (score >= 90) {
|
| 50 |
hintEl.textContent = 'كتابة ممتازة! استمر.';
|
src/model_loader.py
CHANGED
|
@@ -7,16 +7,14 @@ import os
|
|
| 7 |
import logging
|
| 8 |
from pathlib import Path
|
| 9 |
import json
|
| 10 |
-
import pickle
|
| 11 |
import difflib
|
| 12 |
import torch
|
| 13 |
import importlib.util
|
| 14 |
from transformers import (
|
| 15 |
-
MBartForConditionalGeneration,
|
| 16 |
-
AutoTokenizer,
|
| 17 |
AutoConfig,
|
| 18 |
AutoModelForSeq2SeqLM,
|
| 19 |
-
AutoModelForCausalLM,
|
| 20 |
EncoderDecoderModel,
|
| 21 |
BertConfig,
|
| 22 |
EncoderDecoderConfig
|
|
@@ -33,7 +31,6 @@ MODEL_BASE_PATH = Path(__file__).parent.parent / "models"
|
|
| 33 |
SUMMARIZATION_PATH = MODEL_BASE_PATH / "Summarization" / "Model"
|
| 34 |
SPELLING_PATH = MODEL_BASE_PATH / "Spelling" / "Model"
|
| 35 |
AUTOCOMPLETE_PATH = MODEL_BASE_PATH / "Autocomplete" / "Model"
|
| 36 |
-
GRAMMAR_PATH = MODEL_BASE_PATH / "Grammrar" / "Model"
|
| 37 |
PUNCTUATION_PATH = MODEL_BASE_PATH / "Punctuation" / "Model"
|
| 38 |
|
| 39 |
|
|
@@ -91,43 +88,29 @@ class SummarizationModel:
|
|
| 91 |
logger.info(f"Model path validated: {self.model_path}")
|
| 92 |
|
| 93 |
def _fix_generation_config(self):
|
| 94 |
-
"""
|
|
|
|
| 95 |
if self._is_remote_source:
|
| 96 |
return
|
| 97 |
|
| 98 |
-
gen_config_path = self.model_path / "generation_config.json"
|
| 99 |
config_path = self.model_path / "config.json"
|
| 100 |
-
|
|
|
|
| 101 |
try:
|
| 102 |
-
# Fix config.json first (this is the main issue)
|
| 103 |
if config_path.exists():
|
| 104 |
with open(config_path, 'r', encoding='utf-8') as f:
|
| 105 |
config = json.load(f)
|
| 106 |
-
|
| 107 |
-
# Check if early_stopping exists in config and is None
|
| 108 |
if 'early_stopping' in config and config['early_stopping'] is None:
|
| 109 |
-
logger.info("
|
| 110 |
-
|
| 111 |
-
with open(config_path, 'w', encoding='utf-8') as f:
|
| 112 |
-
json.dump(config, f, indent=2, ensure_ascii=False)
|
| 113 |
-
logger.info("Fixed config.json - set early_stopping to True")
|
| 114 |
-
|
| 115 |
-
# Fix generation_config.json
|
| 116 |
if gen_config_path.exists():
|
| 117 |
with open(gen_config_path, 'r', encoding='utf-8') as f:
|
| 118 |
gen_config = json.load(f)
|
| 119 |
-
|
| 120 |
-
# Fix early_stopping if it's None/null
|
| 121 |
if gen_config.get('early_stopping') is None:
|
| 122 |
-
logger.info("
|
| 123 |
-
|
| 124 |
-
with open(gen_config_path, 'w', encoding='utf-8') as f:
|
| 125 |
-
json.dump(gen_config, f, indent=2, ensure_ascii=False)
|
| 126 |
-
logger.info("Fixed generation_config.json - set early_stopping to True")
|
| 127 |
-
|
| 128 |
except Exception as e:
|
| 129 |
-
logger.warning(f"Could not
|
| 130 |
-
# Continue anyway, we'll try to load with workaround
|
| 131 |
|
| 132 |
def _load_model(self):
|
| 133 |
"""Load the model and tokenizer."""
|
|
@@ -145,14 +128,14 @@ class SummarizationModel:
|
|
| 145 |
try:
|
| 146 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 147 |
self.model_source,
|
| 148 |
-
local_files_only=
|
| 149 |
trust_remote_code=False
|
| 150 |
)
|
| 151 |
logger.info("Tokenizer loaded successfully")
|
| 152 |
except Exception as e:
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 157 |
self.model_source,
|
| 158 |
trust_remote_code=False
|
|
@@ -169,7 +152,7 @@ class SummarizationModel:
|
|
| 169 |
try:
|
| 170 |
config = AutoConfig.from_pretrained(
|
| 171 |
self.model_source,
|
| 172 |
-
local_files_only=
|
| 173 |
trust_remote_code=False
|
| 174 |
)
|
| 175 |
|
|
@@ -384,20 +367,27 @@ class SummarizationModel:
|
|
| 384 |
|
| 385 |
class SpellingModel:
|
| 386 |
"""Wrapper class for the Arabic spelling correction model."""
|
| 387 |
-
|
| 388 |
def __init__(self, model_path=None):
|
| 389 |
"""
|
| 390 |
Initialize the spelling model.
|
| 391 |
-
|
| 392 |
Args:
|
| 393 |
model_path: Path to the model directory (defaults to SPELLING_PATH)
|
| 394 |
"""
|
| 395 |
self.model_path = Path(model_path) if model_path else SPELLING_PATH
|
| 396 |
self.model = None
|
| 397 |
self.device = None
|
| 398 |
-
|
| 399 |
self._validate_path()
|
| 400 |
self._load_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 401 |
|
| 402 |
def _validate_path(self):
|
| 403 |
"""Validate that the model path exists."""
|
|
@@ -464,7 +454,7 @@ class SpellingModel:
|
|
| 464 |
# Load state dict
|
| 465 |
pt_file = list(self.model_path.glob("*.pt"))[0]
|
| 466 |
logger.info(f"Loading weights from {pt_file}...")
|
| 467 |
-
checkpoint = torch.load(pt_file, map_location=self.device)
|
| 468 |
|
| 469 |
if "model_state_dict" in checkpoint:
|
| 470 |
state_dict = checkpoint["model_state_dict"]
|
|
@@ -657,161 +647,13 @@ class AutocompleteModel:
|
|
| 657 |
raise RuntimeError(f"Autocomplete prediction failed: {str(e)}")
|
| 658 |
|
| 659 |
|
| 660 |
-
class GrammarModel:
|
| 661 |
-
"""Wrapper class for the Arabic grammar correction model."""
|
| 662 |
-
|
| 663 |
-
def __init__(self, model_path=None):
|
| 664 |
-
"""
|
| 665 |
-
Initialize the grammar model.
|
| 666 |
-
|
| 667 |
-
Args:
|
| 668 |
-
model_path: Path to the model directory (defaults to GRAMMAR_PATH)
|
| 669 |
-
"""
|
| 670 |
-
self.model_path = Path(model_path) if model_path else GRAMMAR_PATH
|
| 671 |
-
self.model = None
|
| 672 |
-
self.tokenizer = None
|
| 673 |
-
self.device = None
|
| 674 |
-
|
| 675 |
-
self._validate_path()
|
| 676 |
-
self._load_model()
|
| 677 |
-
|
| 678 |
-
def _validate_path(self):
|
| 679 |
-
"""Validate that the model path exists and contains required files."""
|
| 680 |
-
if not self.model_path.exists():
|
| 681 |
-
raise FileNotFoundError(f"Model path does not exist: {self.model_path}")
|
| 682 |
-
|
| 683 |
-
required_files = ['config.json', 'tokenizer.json', 'model.safetensors']
|
| 684 |
-
missing_files = [f for f in required_files if not (self.model_path / f).exists()]
|
| 685 |
-
|
| 686 |
-
if missing_files:
|
| 687 |
-
raise FileNotFoundError(f"Missing required files: {', '.join(missing_files)}")
|
| 688 |
-
|
| 689 |
-
logger.info(f"Grammar model path validated: {self.model_path}")
|
| 690 |
-
|
| 691 |
-
def _load_model(self):
|
| 692 |
-
"""Load the grammar model and tokenizer."""
|
| 693 |
-
try:
|
| 694 |
-
# Force CPU-only to avoid GPU OOM / system freezes
|
| 695 |
-
self.device = torch.device('cpu')
|
| 696 |
-
logger.info("Loading grammar model on CPU (GPU disabled by design)...")
|
| 697 |
-
|
| 698 |
-
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 699 |
-
str(self.model_path),
|
| 700 |
-
local_files_only=True
|
| 701 |
-
)
|
| 702 |
-
|
| 703 |
-
self.model = AutoModelForCausalLM.from_pretrained(
|
| 704 |
-
str(self.model_path),
|
| 705 |
-
local_files_only=True,
|
| 706 |
-
trust_remote_code=True,
|
| 707 |
-
torch_dtype=torch.float32 # safe default for CPU inference
|
| 708 |
-
)
|
| 709 |
-
|
| 710 |
-
self.model.to(self.device)
|
| 711 |
-
self.model.eval()
|
| 712 |
-
|
| 713 |
-
logger.info("Grammar model loaded successfully")
|
| 714 |
-
except Exception as e:
|
| 715 |
-
logger.error(f"Error loading grammar model: {str(e)}")
|
| 716 |
-
raise RuntimeError(f"Failed to load grammar model: {str(e)}")
|
| 717 |
-
|
| 718 |
-
def correct(self, text):
|
| 719 |
-
"""
|
| 720 |
-
Correct grammar in Arabic text with a timeout.
|
| 721 |
-
|
| 722 |
-
Args:
|
| 723 |
-
text: Input Arabic text
|
| 724 |
-
|
| 725 |
-
Returns:
|
| 726 |
-
str: Grammar-corrected text
|
| 727 |
-
"""
|
| 728 |
-
if self.model is None or self.tokenizer is None:
|
| 729 |
-
raise RuntimeError("Model not loaded")
|
| 730 |
-
|
| 731 |
-
import threading
|
| 732 |
-
result = [text] # default to original
|
| 733 |
-
error = [None]
|
| 734 |
-
|
| 735 |
-
def _generate():
|
| 736 |
-
try:
|
| 737 |
-
# Use Gemma 3 chat template — pass text only (model is GEC-trained)
|
| 738 |
-
messages = [{"role": "user", "content": text}]
|
| 739 |
-
prompt = self.tokenizer.apply_chat_template(
|
| 740 |
-
messages,
|
| 741 |
-
tokenize=False,
|
| 742 |
-
add_generation_prompt=True,
|
| 743 |
-
)
|
| 744 |
-
|
| 745 |
-
inputs = self.tokenizer(
|
| 746 |
-
prompt,
|
| 747 |
-
max_length=256,
|
| 748 |
-
truncation=True,
|
| 749 |
-
padding=True,
|
| 750 |
-
return_tensors="pt",
|
| 751 |
-
add_special_tokens=False,
|
| 752 |
-
)
|
| 753 |
-
|
| 754 |
-
inputs = {k: v.to(self.device) for k, v in inputs.items()}
|
| 755 |
-
input_length = inputs['input_ids'].shape[1]
|
| 756 |
-
|
| 757 |
-
with torch.no_grad():
|
| 758 |
-
outputs = self.model.generate(
|
| 759 |
-
**inputs,
|
| 760 |
-
max_new_tokens=64, # enough for corrected sentence
|
| 761 |
-
do_sample=False,
|
| 762 |
-
)
|
| 763 |
-
|
| 764 |
-
# Decode only the NEW tokens (skip the prompt)
|
| 765 |
-
new_tokens = outputs[0][input_length:]
|
| 766 |
-
corrected = self.tokenizer.decode(new_tokens, skip_special_tokens=True).strip()
|
| 767 |
-
|
| 768 |
-
# If model returned empty or nonsense, keep original
|
| 769 |
-
if not corrected or len(corrected) < 2:
|
| 770 |
-
return
|
| 771 |
-
|
| 772 |
-
# Reject generic instruction phrases (model fallback, not actual correction)
|
| 773 |
-
generic_phrases = (
|
| 774 |
-
"أعد كتابتها", "أعد كتابته", "أعد كتابتها.", "أعد كتابته.",
|
| 775 |
-
"اعيد كتابتها", "اعيد كتابته", "أعد كتابة", "اعيد كتابة",
|
| 776 |
-
"أعد كتابتها فقط", "أعد كتابته فقط",
|
| 777 |
-
)
|
| 778 |
-
corrected_lower = corrected.strip()
|
| 779 |
-
for phrase in generic_phrases:
|
| 780 |
-
if phrase in corrected_lower or corrected_lower.startswith(phrase):
|
| 781 |
-
return
|
| 782 |
-
|
| 783 |
-
# Take only the first non-empty line as the corrected sentence
|
| 784 |
-
corrected_lines = [l.strip() for l in corrected.split('\n') if l.strip()]
|
| 785 |
-
if corrected_lines:
|
| 786 |
-
first = corrected_lines[0]
|
| 787 |
-
if first and len(first) >= 2 and first not in generic_phrases:
|
| 788 |
-
result[0] = first
|
| 789 |
-
except Exception as e:
|
| 790 |
-
error[0] = e
|
| 791 |
-
|
| 792 |
-
# Run generation in a thread with a 30s timeout
|
| 793 |
-
thread = threading.Thread(target=_generate)
|
| 794 |
-
thread.start()
|
| 795 |
-
thread.join(timeout=30)
|
| 796 |
-
|
| 797 |
-
if thread.is_alive():
|
| 798 |
-
logger.warning("[Grammar] Timed out after 30s — returning original text")
|
| 799 |
-
return text
|
| 800 |
-
|
| 801 |
-
if error[0]:
|
| 802 |
-
logger.error(f"Error during grammar correction: {str(error[0])}")
|
| 803 |
-
logger.warning("Returning original text due to grammar error.")
|
| 804 |
-
|
| 805 |
-
return result[0]
|
| 806 |
-
|
| 807 |
-
|
| 808 |
class PunctuationModel:
|
| 809 |
"""Wrapper class for the Arabic punctuation model."""
|
| 810 |
-
|
| 811 |
def __init__(self, model_path=None):
|
| 812 |
"""
|
| 813 |
Initialize the punctuation model.
|
| 814 |
-
|
| 815 |
Args:
|
| 816 |
model_path: Path to the model directory (defaults to PUNCTUATION_PATH)
|
| 817 |
"""
|
|
@@ -819,9 +661,19 @@ class PunctuationModel:
|
|
| 819 |
self.model = None
|
| 820 |
self.tokenizer = None
|
| 821 |
self.device = None
|
| 822 |
-
|
| 823 |
self._validate_path()
|
| 824 |
self._load_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 825 |
|
| 826 |
def _validate_path(self):
|
| 827 |
"""Validate that the model path exists and contains required files."""
|
|
|
|
| 7 |
import logging
|
| 8 |
from pathlib import Path
|
| 9 |
import json
|
|
|
|
| 10 |
import difflib
|
| 11 |
import torch
|
| 12 |
import importlib.util
|
| 13 |
from transformers import (
|
| 14 |
+
MBartForConditionalGeneration,
|
| 15 |
+
AutoTokenizer,
|
| 16 |
AutoConfig,
|
| 17 |
AutoModelForSeq2SeqLM,
|
|
|
|
| 18 |
EncoderDecoderModel,
|
| 19 |
BertConfig,
|
| 20 |
EncoderDecoderConfig
|
|
|
|
| 31 |
SUMMARIZATION_PATH = MODEL_BASE_PATH / "Summarization" / "Model"
|
| 32 |
SPELLING_PATH = MODEL_BASE_PATH / "Spelling" / "Model"
|
| 33 |
AUTOCOMPLETE_PATH = MODEL_BASE_PATH / "Autocomplete" / "Model"
|
|
|
|
| 34 |
PUNCTUATION_PATH = MODEL_BASE_PATH / "Punctuation" / "Model"
|
| 35 |
|
| 36 |
|
|
|
|
| 88 |
logger.info(f"Model path validated: {self.model_path}")
|
| 89 |
|
| 90 |
def _fix_generation_config(self):
|
| 91 |
+
"""Check generation configs for early_stopping=None and log a warning.
|
| 92 |
+
In-memory fix is applied after loading at line ~177."""
|
| 93 |
if self._is_remote_source:
|
| 94 |
return
|
| 95 |
|
|
|
|
| 96 |
config_path = self.model_path / "config.json"
|
| 97 |
+
gen_config_path = self.model_path / "generation_config.json"
|
| 98 |
+
|
| 99 |
try:
|
|
|
|
| 100 |
if config_path.exists():
|
| 101 |
with open(config_path, 'r', encoding='utf-8') as f:
|
| 102 |
config = json.load(f)
|
|
|
|
|
|
|
| 103 |
if 'early_stopping' in config and config['early_stopping'] is None:
|
| 104 |
+
logger.info("early_stopping is None in config.json — will fix in-memory after load")
|
| 105 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
if gen_config_path.exists():
|
| 107 |
with open(gen_config_path, 'r', encoding='utf-8') as f:
|
| 108 |
gen_config = json.load(f)
|
|
|
|
|
|
|
| 109 |
if gen_config.get('early_stopping') is None:
|
| 110 |
+
logger.info("early_stopping is None in generation_config.json — will fix in-memory after load")
|
| 111 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
except Exception as e:
|
| 113 |
+
logger.warning(f"Could not read generation config files: {str(e)}")
|
|
|
|
| 114 |
|
| 115 |
def _load_model(self):
|
| 116 |
"""Load the model and tokenizer."""
|
|
|
|
| 128 |
try:
|
| 129 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 130 |
self.model_source,
|
| 131 |
+
local_files_only=not self._is_remote_source,
|
| 132 |
trust_remote_code=False
|
| 133 |
)
|
| 134 |
logger.info("Tokenizer loaded successfully")
|
| 135 |
except Exception as e:
|
| 136 |
+
if not self._is_remote_source:
|
| 137 |
+
logger.error(f"Failed to load tokenizer: {str(e)}")
|
| 138 |
+
logger.info("Retrying tokenizer load without local_files_only...")
|
| 139 |
self.tokenizer = AutoTokenizer.from_pretrained(
|
| 140 |
self.model_source,
|
| 141 |
trust_remote_code=False
|
|
|
|
| 152 |
try:
|
| 153 |
config = AutoConfig.from_pretrained(
|
| 154 |
self.model_source,
|
| 155 |
+
local_files_only=not self._is_remote_source,
|
| 156 |
trust_remote_code=False
|
| 157 |
)
|
| 158 |
|
|
|
|
| 367 |
|
| 368 |
class SpellingModel:
|
| 369 |
"""Wrapper class for the Arabic spelling correction model."""
|
| 370 |
+
|
| 371 |
def __init__(self, model_path=None):
|
| 372 |
"""
|
| 373 |
Initialize the spelling model.
|
| 374 |
+
|
| 375 |
Args:
|
| 376 |
model_path: Path to the model directory (defaults to SPELLING_PATH)
|
| 377 |
"""
|
| 378 |
self.model_path = Path(model_path) if model_path else SPELLING_PATH
|
| 379 |
self.model = None
|
| 380 |
self.device = None
|
| 381 |
+
|
| 382 |
self._validate_path()
|
| 383 |
self._load_model()
|
| 384 |
+
|
| 385 |
+
def __del__(self):
|
| 386 |
+
if self.model is not None:
|
| 387 |
+
del self.model
|
| 388 |
+
self.model = None
|
| 389 |
+
if torch.cuda.is_available():
|
| 390 |
+
torch.cuda.empty_cache()
|
| 391 |
|
| 392 |
def _validate_path(self):
|
| 393 |
"""Validate that the model path exists."""
|
|
|
|
| 454 |
# Load state dict
|
| 455 |
pt_file = list(self.model_path.glob("*.pt"))[0]
|
| 456 |
logger.info(f"Loading weights from {pt_file}...")
|
| 457 |
+
checkpoint = torch.load(pt_file, map_location=self.device, weights_only=True)
|
| 458 |
|
| 459 |
if "model_state_dict" in checkpoint:
|
| 460 |
state_dict = checkpoint["model_state_dict"]
|
|
|
|
| 647 |
raise RuntimeError(f"Autocomplete prediction failed: {str(e)}")
|
| 648 |
|
| 649 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 650 |
class PunctuationModel:
|
| 651 |
"""Wrapper class for the Arabic punctuation model."""
|
| 652 |
+
|
| 653 |
def __init__(self, model_path=None):
|
| 654 |
"""
|
| 655 |
Initialize the punctuation model.
|
| 656 |
+
|
| 657 |
Args:
|
| 658 |
model_path: Path to the model directory (defaults to PUNCTUATION_PATH)
|
| 659 |
"""
|
|
|
|
| 661 |
self.model = None
|
| 662 |
self.tokenizer = None
|
| 663 |
self.device = None
|
| 664 |
+
|
| 665 |
self._validate_path()
|
| 666 |
self._load_model()
|
| 667 |
+
|
| 668 |
+
def __del__(self):
|
| 669 |
+
if self.model is not None:
|
| 670 |
+
del self.model
|
| 671 |
+
self.model = None
|
| 672 |
+
if self.tokenizer is not None:
|
| 673 |
+
del self.tokenizer
|
| 674 |
+
self.tokenizer = None
|
| 675 |
+
if torch.cuda.is_available():
|
| 676 |
+
torch.cuda.empty_cache()
|
| 677 |
|
| 678 |
def _validate_path(self):
|
| 679 |
"""Validate that the model path exists and contains required files."""
|
src/nlp/spelling/contextual_corrector.py
CHANGED
|
@@ -12,6 +12,7 @@
|
|
| 12 |
import logging
|
| 13 |
import torch
|
| 14 |
import threading
|
|
|
|
| 15 |
from typing import List, Tuple, Optional, Dict
|
| 16 |
|
| 17 |
logger = logging.getLogger(__name__)
|
|
@@ -19,7 +20,7 @@ logger = logging.getLogger(__name__)
|
|
| 19 |
# Singleton instance
|
| 20 |
_instance = None
|
| 21 |
_loading = False
|
| 22 |
-
_lock = threading.
|
| 23 |
|
| 24 |
|
| 25 |
class ContextualCorrector:
|
|
@@ -44,7 +45,7 @@ class ContextualCorrector:
|
|
| 44 |
self.model.eval()
|
| 45 |
|
| 46 |
# Simple cache for scores
|
| 47 |
-
self._cache:
|
| 48 |
self._cache_max = 5000
|
| 49 |
|
| 50 |
# Vocab for candidate filtering
|
|
@@ -63,8 +64,9 @@ class ContextualCorrector:
|
|
| 63 |
Returns:
|
| 64 |
Probability score (0.0 to 1.0) — higher = better fit
|
| 65 |
"""
|
| 66 |
-
cache_key = f"{
|
| 67 |
if cache_key in self._cache:
|
|
|
|
| 68 |
return self._cache[cache_key]
|
| 69 |
|
| 70 |
words = text.split()
|
|
@@ -99,13 +101,10 @@ class ContextualCorrector:
|
|
| 99 |
if not word_tokens:
|
| 100 |
return 0.0
|
| 101 |
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
for t_id in word_tokens[1:]:
|
| 107 |
-
score *= probs[t_id].item()
|
| 108 |
-
score = score ** (1.0 / len(word_tokens))
|
| 109 |
|
| 110 |
except Exception as e:
|
| 111 |
logger.warning(f"[MLM] Score error for '{word}': {e}")
|
|
@@ -113,10 +112,8 @@ class ContextualCorrector:
|
|
| 113 |
|
| 114 |
# Cache management
|
| 115 |
if len(self._cache) >= self._cache_max:
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
for k in keys_to_remove:
|
| 119 |
-
del self._cache[k]
|
| 120 |
self._cache[cache_key] = score
|
| 121 |
|
| 122 |
return score
|
|
@@ -175,8 +172,8 @@ class ContextualCorrector:
|
|
| 175 |
if corr_score > confidence_threshold:
|
| 176 |
continue
|
| 177 |
|
| 178 |
-
# Score the original word in
|
| 179 |
-
orig_score = self.score_word_in_context(
|
| 180 |
|
| 181 |
# If original scores better, revert
|
| 182 |
if orig_score > corr_score * 10 and orig_score > 0.01:
|
|
@@ -271,9 +268,6 @@ class ContextualCorrector:
|
|
| 271 |
if not a or not b:
|
| 272 |
return 0.0
|
| 273 |
max_len = max(len(a), len(b))
|
| 274 |
-
if max_len == 0:
|
| 275 |
-
return 1.0
|
| 276 |
-
# Inline Levenshtein to avoid extra dependency
|
| 277 |
m, n = len(a), len(b)
|
| 278 |
dp = list(range(n + 1))
|
| 279 |
for i in range(1, m + 1):
|
|
|
|
| 12 |
import logging
|
| 13 |
import torch
|
| 14 |
import threading
|
| 15 |
+
from collections import OrderedDict
|
| 16 |
from typing import List, Tuple, Optional, Dict
|
| 17 |
|
| 18 |
logger = logging.getLogger(__name__)
|
|
|
|
| 20 |
# Singleton instance
|
| 21 |
_instance = None
|
| 22 |
_loading = False
|
| 23 |
+
_lock = threading.RLock()
|
| 24 |
|
| 25 |
|
| 26 |
class ContextualCorrector:
|
|
|
|
| 45 |
self.model.eval()
|
| 46 |
|
| 47 |
# Simple cache for scores
|
| 48 |
+
self._cache: OrderedDict[str, float] = OrderedDict()
|
| 49 |
self._cache_max = 5000
|
| 50 |
|
| 51 |
# Vocab for candidate filtering
|
|
|
|
| 64 |
Returns:
|
| 65 |
Probability score (0.0 to 1.0) — higher = better fit
|
| 66 |
"""
|
| 67 |
+
cache_key = f"{hash(text)}|{position}|{word}"
|
| 68 |
if cache_key in self._cache:
|
| 69 |
+
self._cache.move_to_end(cache_key)
|
| 70 |
return self._cache[cache_key]
|
| 71 |
|
| 72 |
words = text.split()
|
|
|
|
| 101 |
if not word_tokens:
|
| 102 |
return 0.0
|
| 103 |
|
| 104 |
+
# Use first subword token probability only. For multi-subword
|
| 105 |
+
# words, the single [MASK] distribution only predicts the first
|
| 106 |
+
# subword — multiplying unrelated token probs is nonsensical.
|
| 107 |
+
score = probs[word_tokens[0]].item()
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
except Exception as e:
|
| 110 |
logger.warning(f"[MLM] Score error for '{word}': {e}")
|
|
|
|
| 112 |
|
| 113 |
# Cache management
|
| 114 |
if len(self._cache) >= self._cache_max:
|
| 115 |
+
for _ in range(self._cache_max // 5):
|
| 116 |
+
self._cache.popitem(last=False)
|
|
|
|
|
|
|
| 117 |
self._cache[cache_key] = score
|
| 118 |
|
| 119 |
return score
|
|
|
|
| 172 |
if corr_score > confidence_threshold:
|
| 173 |
continue
|
| 174 |
|
| 175 |
+
# Score the original word in its own context (not contaminated by other corrections)
|
| 176 |
+
orig_score = self.score_word_in_context(original_text, i, orig_w)
|
| 177 |
|
| 178 |
# If original scores better, revert
|
| 179 |
if orig_score > corr_score * 10 and orig_score > 0.01:
|
|
|
|
| 268 |
if not a or not b:
|
| 269 |
return 0.0
|
| 270 |
max_len = max(len(a), len(b))
|
|
|
|
|
|
|
|
|
|
| 271 |
m, n = len(a), len(b)
|
| 272 |
dp = list(range(n + 1))
|
| 273 |
for i in range(1, m + 1):
|
src/nlp/stage_locker.py
CHANGED
|
@@ -94,6 +94,8 @@ class StageLocker:
|
|
| 94 |
req_priority = STAGE_PRIORITY.get(requesting_stage, 0)
|
| 95 |
for ls, le, owner in self.locked_spans:
|
| 96 |
if start < le and end > ls:
|
|
|
|
|
|
|
| 97 |
owner_priority = STAGE_PRIORITY.get(owner, 0)
|
| 98 |
if owner_priority >= req_priority:
|
| 99 |
if PIPELINE_DEBUG:
|
|
@@ -161,6 +163,8 @@ class StageLocker:
|
|
| 161 |
req_priority = STAGE_PRIORITY.get(requesting_stage, 0)
|
| 162 |
for ls, le, owner in self.locked_spans:
|
| 163 |
if start < le and end > ls:
|
|
|
|
|
|
|
| 164 |
owner_priority = STAGE_PRIORITY.get(owner, 0)
|
| 165 |
if owner_priority >= req_priority:
|
| 166 |
return (ls, le, owner)
|
|
|
|
| 94 |
req_priority = STAGE_PRIORITY.get(requesting_stage, 0)
|
| 95 |
for ls, le, owner in self.locked_spans:
|
| 96 |
if start < le and end > ls:
|
| 97 |
+
if owner == 'protection':
|
| 98 |
+
return True
|
| 99 |
owner_priority = STAGE_PRIORITY.get(owner, 0)
|
| 100 |
if owner_priority >= req_priority:
|
| 101 |
if PIPELINE_DEBUG:
|
|
|
|
| 163 |
req_priority = STAGE_PRIORITY.get(requesting_stage, 0)
|
| 164 |
for ls, le, owner in self.locked_spans:
|
| 165 |
if start < le and end > ls:
|
| 166 |
+
if owner == 'protection':
|
| 167 |
+
return (ls, le, owner)
|
| 168 |
owner_priority = STAGE_PRIORITY.get(owner, 0)
|
| 169 |
if owner_priority >= req_priority:
|
| 170 |
return (ls, le, owner)
|