| """Generate the SpellBench training and test datasets. |
| Produces items per task across word-level and sentence-level categories. |
| Tracks language and script for each sample. |
| |
| Usage: |
| python generate.py |
| """ |
| import json, os, re, random, string, argparse |
| from pathlib import Path |
| from collections import Counter |
|
|
| |
|
|
| |
| WORD_POOLS = [ |
| |
| ("Latin", "English", [ |
| "hello", "world", "butterfly", "elephant", "strawberry", "computer", "language", "beautiful", "mountain", "adventure", |
| "chocolate", "happiness", "umbrella", "keyboard", "telescope", "caterpillar", "octopus", "labyrinth", "algorithm", "xylophone", |
| "quartz", "rhythm", "zephyr", "sphinx", "lynx", "fjord", "glyph", "knack", "jazz", "fizz", "buzz", "queue", "aqua", "kayak", |
| "level", "radar", "rotor", "madam", "civic", "refer", "deified", "racecar", "noon", "deed", "peep", "toot", "pup", "mom", "dad", |
| "eye", "sis", "nun", "tot", "bib", "gig", "pip", "pop", "banana", "orange", "cherry", "mango", "pineapple", "avocado", |
| "watermelon", "blueberry", "raspberry", "blackberry", "coconut", "apricot", "pomegranate", "cantaloupe", "tangerine", "persimmon", |
| "abstraction", "benevolent", "catastrophe", "dilapidated", "ephemeral", "fortuitous", "gregarious", "hypothetical", "idiosyncratic", |
| "juxtapose", "kaleidoscope", "loquacious", "metamorphosis", "nonchalant", "obsequious", "perpendicular", "quintessential", |
| "recalcitrant", "serendipity", "transcendent", "ubiquitous", "verisimilitude", "whimsical", "xenophobia", "yearning", "zealous", |
| "abandon", "ability", "absence", "absolute", "academic", "accept", "accident", "accompany", "accurate", "achieve" |
| ]), |
| ("Latin", "French", [ |
| "bonjour", "merci", "maison", "fenêtre", "château", "bibliothèque", "papillon", "éléphant", "croissant", "baguette", "résumé", |
| "naïve", "café", "cliché", "fiancée", "soirée", "début", "décor", "détour", "ordinateur", "voyage", "liberté", "égalité", |
| "fraternité", "étoile", "montagne", "rivière", "océan", "soleil", "lune", "fleur", "pomme", "livre", "oiseau", "chien", "chat", |
| "ami", "amour", "arbre", "argent", "avion", "bateau", "blanc", "bleu", "bonheur", "bougie" |
| ]), |
| ("Latin", "German", [ |
| "schmetterling", "kindergarten", "wanderlust", "zeitgeist", "angst", "doppelgänger", "über", "gemütlichkeit", "schadenfreude", |
| "hinterland", "blitzkrieg", "poltergeist", "rucksack", "wunderbar", "strafe", "geschwindigkeitsbegrenzung", "nachricht", "zug", |
| "flughafen", "bahnhof", "kirche", "straße", "brücke", "wald", "meer", "regen", "schnee", "sonne", "mond", "himmel", |
| "apfel", "buch", "danke", "essen", "freund", "geld", "haus", "immer", "jetzt", "kalt" |
| ]), |
| ("Latin", "Spanish", [ |
| "mariposa", "corazón", "biblioteca", "murciélago", "escarabajo", "amanecer", "esperanza", "madrugada", "desayuno", "almohada", |
| "ventana", "puerta", "camino", "montaña", "bosque", "playa", "estrella", "computadora", "teléfono", "idioma", "amistad", |
| "agua", "azul", "blanco", "calle", "ciudad", "comida", "dormir", "escuela", "familia", "fuego" |
| ]), |
| ("Latin", "Indonesian", [ |
| "selamat", "terima", "kasih", "perpustakaan", "matahari", "rembulan", "bintang", "gunung", "pantai", "pelangi", "komputer", |
| "bahasa", "merdeka", "pancasila", "indonesia", "keluarga", "sahabat", "belajar", "sekolah", "gambar", "cerita", |
| "air", "api", "buku", "dunia", "guru", "hari", "ikan", "jalan", "makan", "minum" |
| ]), |
| ("Latin", "Swahili", [ |
| "habari", "asante", "maktaba", "nyumbani", "jua", "mwezi", "nyota", "mlima", "bahari", "chemchemi", "kompyuta", "lugha", |
| "uhuru", "amani", "upendo", "rafiki", "familia", "shule", "kitabu", "maua", "mti", |
| "maji", "chakula", "safari", "habari", "ndiyo", "hapana", "vizuri", "leo", "kesho", "sasa" |
| ]), |
| ("Latin", "Yoruba", [ |
| "ẹ kàárọ̀", "ẹ kú ṣẹ́", "ilé-ìwé", "àyà", "ọ̀rẹ́", "ìfẹ́", "ìrètí", "ìgbàgbọ́", "àyọ̀", "àlàáfíà", "àṣà", "ẹ̀kọ́", "orílẹ̀-èdè", |
| "ẹ̀mí", "ọpẹ́", "ojú", "itàn", "ewì", "orin", "idùnnú", |
| "omi", "onje", "baba", "iyá", "omo", "ilé", "ọ̀nà", "ọ̀run", "ayé", "ọkàn" |
| ]), |
| |
| ("Latin", "Moroccan Arabizi", [ |
| "daba", "salam", "labas", "choukran", "ghadi", "3ziz", "fine", "mcha", "khlass", "dar", "bit", "madda", "me39ol", "khalina", "wakha", "beslama", |
| "7chouma", "9raya", "mabrouk", "inchallah", "safi", "drari", "mza7", "mra", "rajel", "macha", "mchi", "douz", "me7ro9a", |
| "homa", "ntoma", "hna", "liyoum", "ghda", "frain", "toubis", "triq", "wad", "bhar" |
| ]), |
| ("Latin", "Egyptian Arabizi", [ |
| "ezayak", "amr", "kwayes", "ya3ni", "delwa2ty", "bokra", "enaharda", "shokran", "7aga", "moshkela", "begad", "3ashan", "leih", "mashi", "yalla", |
| "habibi", "raya7", "geh", "fayn", "7elw", "shwaya", "keteer", "ayer", "bas", "3reft", "7abit", "kont", "oltelak", "7ad", "kol", "da", "howa", |
| "esta3mel", "mat3am", "shari3", "beit", "maia", "akl", "ishta", "gamed", "fashkh", "ya" |
| ]), |
| |
| ("Latin", "Romanized Russian", [ |
| "babochka", "privet", "spasibo", "krasivyj", "biblioteka", "kompyuter", "dozhd", "solnce", "zvezda", "moloko", "zdravstvujte", |
| "pozhalujsta", "izvinite", "horosho", "bolshoj", "malenkij", "chelovek", "druzhba", "lyubov", "gorod", "doroga", "mashina", |
| "voda", "eda", "kniga", "shkola", "rabota", "utro", "vecher", "noch", "den", "vse" |
| ]), |
| ("Latin", "Romanized Japanese", [ |
| "konnichiwa", "arigatou", "sayounara", "sumimasen", "ohayou", "konbanwa", "oyasumi", "itadakimasu", "gochisousama", "hajimemashite", |
| "yoroshiku", "onegaishimasu", "tadaima", "okaeri", "ittekimasu", "itterasshai", "omedetou", "ganbatte", "ki o tsukete", "daijoubu", |
| "mizu", "tabemono", "hon", "gakkou", "shigoto", "asa", "yoru", "hi", "tsuki", "hana" |
| ]), |
| |
| ("Arabic", "Arabic", [ |
| "مكتبة", "كتاب", "مدرسة", "حاسوب", "شمس", "قمر", "بحر", "جبل", "سماء", "أرض", "ماء", "نار", "هواء", "شجرة", "زهرة", "فراشة", |
| "عصفور", "حصان", "سمكة", "قطة", "كلب", "فيل", "أسد", "نمر", "طائرة", "سيارة", "قطار", "باب", "نافذة", "مفتاح", "حضارة", |
| "تاريخ", "ثقافة", "فن", "علم", "نور", "ظلام", "سلام", "حب", "سعادة", "جمال", "صوت", "كلمة", |
| "صديق", "حياة", "امل", "عمل", "بيت", "طريق", "خريطة", "بلد", "شعب", "لغة" |
| ]), |
| ("Arabic", "Arabic (Tashkeel)", [ |
| "مَكْتَبَةٌ", "كِتَابٌ", "مَدْرَسَةٌ", "حَاسُوبٌ", "شَمْسٌ", "قَمَرٌ", "بَحْرٌ", "جَبَلٌ", "سَمَاءٌ", "أَرْضٌ", "مَاءٌ", "نَارٌ", |
| "هَوَاءٌ", "شَجَرَةٌ", "زَهْرَةٌ", "فَرَاشَةٌ", "عُصْفُورٌ", "حِصَانٌ", "سَمَكَةٌ", "قِطَّةٌ", "كَلْبٌ", "فِيلٌ", "أَسَدٌ", "نَمِرٌ", |
| "طَائِرَةٌ", "سَيَّارَةٌ", "قِطَارٌ", "بَابٌ", "نَافِذَةٌ", "مِفْتَاحٌ", "حَضَارَةٌ", "تَارِيخٌ", "ثَقَافَةٌ", "فَنٌّ", "عِلْمٌ", |
| "نُورٌ", "ظَلَامٌ", "سَلَامٌ", "حُبٌّ", "سَعَادَةٌ", "جَمَالٌ", "صَوْتٌ", "كَلِمَةٌ", |
| "صَدِيقٌ", "حَيَاةٌ", "أَمَلٌ", "عَمَلٌ", "بَيْتٌ", "طَرِيقٌ", "خَرِيطَةٌ", "بَلَدٌ", "شَعْبٌ", "لُغَةٌ" |
| ]), |
| ("Arabic", "Persian", [ |
| "کتابخانه", "خورشید", "پروانه", "درخت", "آسمان", "زمین", "دوست", "زندگی", "شادی", "ستاره", "دریا", "کوه", "گل", "پرنده", "ماه", |
| "باد", "آتش", "آب", "مرد", "زن", "کودک", "خانواده", "فرهنگ", "عشق", "صلح", "راه", "ماشین", "کامپیوتر", "پنجره", "نور", |
| "کتاب", "مدرسه", "میز", "صندلی", "قلم", "دفتر", "امروز", "فردا", "دیروز", "سلام" |
| ]), |
| |
| ("Cyrillic", "Russian", [ |
| "бабочка", "привет", "спасибо", "красивый", "библиотека", "компьютер", "дождь", "солнце", "звезда", "молоко", "здравствуйте", |
| "пожалуйста", "извините", "хорошо", "большой", "маленький", "человек", "дружба", "любовь", "город", "дорога", "машина", |
| "вода", "еда", "книга", "школа", "работа", "утро", "вечер", "ночь", "день", "все" |
| ]), |
| ("Cyrillic", "Bulgarian", [ |
| "благодаря", "планина", "река", "море", "свобода", "приятел", "котка", "куче", "къща", "прозорец", "врата", "слънце", "луна", |
| "звезда", "небе", "земя", "огън", "вода", "вятър", "дърво", "цвете", "птица", "кон", "кола", "влак", "самолет", "град", "село", |
| "път", "мост", "час", "ден", "нощ", "лято", "зима", "пролет", "есен", "хляб", "сол", "мед" |
| ]), |
| |
| ("Devanagari", "Hindi", [ |
| "नमस्ते", "पुस्तक", "विद्यालय", "कंप्यूटर", "तितली", "सूरज", "चांद", "पानी", "हवा", "फूल", "पेड़", "पहाड़", "नदी", "समुद्र", "आकाश", |
| "धरती", "आदमी", "औरत", "बच्चा", "मित्र", "प्यार", "खुशी", "शांति", "शहर", "सड़क", "गाड़ी", "बि़ल्ली", "अंधेरा", "रोशनी", |
| "देश", "भाषा", "खाना", "सोना", "जाना", "आना", "काम", "खेल", "जीवन", "मौत" |
| ]), |
| ("Devanagari", "Marathi", [ |
| "नमस्कार", "शाळा", "पुस्तक", "संगणक", "फुलपाखरू", "सूर्य", "चंद्र", "पाणी", "वारा", "फूल", "झाड", "डोंगर", "नदी", "आकाश", "पृथ्वी", |
| "माणूस", "स्त्री", "मुलगा", "मुलगी", "मित्र", "प्रेम", "आनंद", "शांतता", "गाव", "शहरी", "रस्ता", |
| "घर", "काम", "जेवण", "पाऊस", "बाजार", "रंग", "वेळ", "दिवस", "रात्र", "शिक्षण" |
| ]), |
| |
| ("Greek", "Greek", [ |
| "γεια", "ευχαριστώ", "βιβλιοθήκη", "αυτοκίνητο", "θάλασσα", "βουνό", "ουρανός", "αστέρι", "ήλιος", "φεγγάρι", "λουλούδι", |
| "δέντρο", "πεταλούδα", "υπολογιστής", "άνθρωπος", "αγάπη", "ελπίδα", "ειρήνη", "κόσμος", "γλώσσα", "ιστορία", "επιστήμη", |
| "νερό", "φαγητό", "σπίτι", "δρόμος", "πόλη", "γραφή", "βιβλίο", "μέρα", "νύχτα" |
| ]), |
| ("Armenian", "Armenian", [ |
| "բարև", "շնորհակալություն", "գրադարան", "մեքենա", "ծով", "լեռ", "երկինք", "աստղ", "արև", "լուսին", "ծաղիկ", "ծառ", "թիթեռ", |
| "համակարգիչ", "մարդ", "սեր", "հույս", "խաղաղություն", "աշխարհ", "լեզու", "քաղաք", |
| "ջուր", "հաց", "տուն", "ճանապարհ", "գիրք", "օր", "գիշեր", "ընկեր", "ընտանիք" |
| ]), |
| ("Georgian", "Georgian", [ |
| "გამარჯობა", "მადლობა", "ბიბლიოთეკა", "მანქანა", "ზღვა", "მთა", "ცის", "ვარსკვლავი", "მზე", "მთვარე", "ყვავილი", "ხე", "პეპელა", |
| "კომპიუტერი", "ადამიანი", "სიყვარული", "იმედი", "მშვიდობა", "სამყარო", "ენა", "ქალაქი", |
| "წყალი", "პური", "სახლი", "გზა", "წიგნი", "დღე", "ღამე", "მეგობარი", "ოჯახი" |
| ]), |
| ("Hangul", "Korean", [ |
| "안녕하세요", "감사합니다", "도서관", "자동차", "바다", "산", "하늘", "별", "해", "달", "꽃", "나무", "나비", |
| "컴퓨터", "사람", "사랑", "희망", "평화", "세계", "언어", "도시", |
| "물", "빵", "집", "길", "책", "날", "밤", "친구", "가족" |
| ]), |
| ("Thai", "Thai", [ |
| "สวัสดี", "ขอบคุณ", "ห้องสมุด", "รถยนต์", "ทะเล", "ภูเขา", "ท้องฟ้า", "ดาว", "พระอาทิตย์", "พระจันทร์", "ดอกไม้", "ต้นไม้", "ผีเสื้อ", |
| "คอมพิวเตอร์", "คน", "รัก", "ความหวัง", "สันติภาพ", "โลก", "ภาษา", "เมือง", |
| "น้ำ", "ข้าว", "บ้าน", "ทาง", "หนังสือ", "วัน", "คืน", "เพื่อน", "ครอบครัว" |
| ]), |
| ("Hebrew", "Hebrew", [ |
| "שלום", "תודה", "ספר", "בית", "שמש", "ירח", "כוכב", "ים", "הר", "שמיים", |
| "ארץ", "מים", "אש", "עץ", "פרח", "פרפר", "ציפור", "סוס", "דג", "חתול", |
| "כלב", "פיל", "אריה", "נמר", "מכונית", "רכבת", "מטוס", "בית-ספר", "מחשב", "טלפון", |
| "שפה", "היסטוריה", "מדע", "מוזיקה", "ספורט", "אוכל", "לחם", "יין", "חבר", "משפחה", |
| "אהבה", "תקווה", "שמחה", "יופי", "אמת", "צדק", "חופש", "לילה", "יום", |
| "בוקר", "ערב", "שנה", "חודש", "שבוע", "זמן", "עולם", "עיר", "כפר", "דרך" |
| ]), |
| ] |
|
|
| VOWELS = set("aeiou") |
|
|
| |
|
|
| def _char_script(ch): |
| """Return the Unicode script family of a character.""" |
| import unicodedata |
| if ch in ' \t-_\'"' or ch.isdigit(): |
| return 'Common' |
| if unicodedata.category(ch) == 'Mn': |
| return 'Common' |
| try: |
| name = unicodedata.name(ch, '').upper() |
| except Exception: |
| return 'Unknown' |
| for label in ['LATIN', 'ARABIC', 'CYRILLIC', 'DEVANAGARI', 'GREEK', |
| 'ARMENIAN', 'GEORGIAN', 'HANGUL', 'THAI', 'HEBREW']: |
| if label in name: |
| return label |
| return f'OTHER({unicodedata.category(ch)})' |
|
|
| _ALLOWED_SCRIPTS = { |
| 'Latin': {'LATIN', 'Common'}, |
| 'Arabic': {'ARABIC', 'Common'}, |
| 'Cyrillic': {'CYRILLIC', 'Common'}, |
| 'Devanagari': {'DEVANAGARI', 'Common'}, |
| 'Greek': {'GREEK', 'Common'}, |
| 'Armenian': {'ARMENIAN', 'Common'}, |
| 'Georgian': {'GEORGIAN', 'Common'}, |
| 'Hangul': {'HANGUL', 'Common'}, |
| 'Thai': {'THAI', 'Common'}, |
| 'Hebrew': {'HEBREW', 'Common'}, |
| } |
|
|
| def validate_pools(pools): |
| """Check every word in every pool for cross-script contamination. |
| Raises ValueError listing all offending words, or returns True if clean.""" |
| issues = [] |
| for exp_script, lang, words in pools: |
| allowed = _ALLOWED_SCRIPTS.get(exp_script, {'Common'}) |
| for word in words: |
| bad = [(ch, _char_script(ch)) for ch in word if _char_script(ch) not in allowed] |
| if bad: |
| issues.append(f' [{lang}] {repr(word)}: unexpected chars {bad}') |
| if issues: |
| raise ValueError('Cross-script contamination detected:\n' + '\n'.join(issues)) |
| return True |
|
|
| def get_balanced_words(words, split_ratio=0.5): |
| """Split words into train and test sets.""" |
| shuffled = words[:] |
| random.shuffle(shuffled) |
| split_idx = int(len(shuffled) * split_ratio) |
| return shuffled[:split_idx], shuffled[split_idx:] |
|
|
| def get_word_samples(pools, per_lang=100): |
| """Draw samples from provided pools.""" |
| samples = [] |
| for script, lang, words in pools: |
| if len(words) < per_lang: |
| lang_samples = random.choices(words, k=per_lang) |
| else: |
| lang_samples = random.sample(words, per_lang) |
| for w in lang_samples: |
| samples.append({"word": w, "language": lang, "script": script}) |
| return samples |
|
|
| def make_sentence(lang_pool, min_words=3, max_words=7): |
| n = random.randint(min_words, max_words) |
| words = random.choices(lang_pool, k=n) |
| return " ".join(words), words |
|
|
| def gen_task_base(task_name, logic_fn, pools, category="word", sentence_logic=False, per_lang=100): |
| items = [] |
| if category == "word": |
| samples = get_word_samples(pools, per_lang=per_lang) |
| for s in samples: |
| w = s["word"] |
| if task_name in ["pig_latin", "vowel_count", "consonant_count", "remove_vowels"] and s["script"] != "Latin": |
| continue |
| expected = logic_fn(w) |
| items.append({ |
| "task": task_name, |
| "input": w, |
| "expected": str(expected), |
| "metadata": {"language": s["language"], "script": s["script"], "word": w} |
| }) |
| elif category == "sentence": |
| for script, lang, words in pools: |
| for _ in range(per_lang): |
| sentence, word_list = make_sentence(words) |
| expected = logic_fn(sentence, word_list) if sentence_logic else logic_fn(sentence) |
| items.append({ |
| "task": task_name, |
| "input": sentence, |
| "expected": str(expected), |
| "metadata": {"language": lang, "script": script, "sentence": sentence} |
| }) |
| return items |
|
|
| |
| def spell(w): return "-".join(w) |
| def reverse(w): return w[::-1] |
| def word_length(w): return len(w) |
| def first_letter(w): return w[0] |
| def last_letter(w): return w[-1] |
| def vowel_count(w): return sum(1 for c in w.lower() if c in VOWELS) |
| def consonant_count(w): return sum(1 for c in w.lower() if c.isalpha() and c not in VOWELS) |
| def remove_vowels(w): return "".join(c for c in w if c.lower() not in VOWELS) |
| def is_palindrome(w): return str(w.lower() == w.lower()[::-1]).lower() |
| def word_count(s, wl): return len(wl) |
| def sentence_reverse(s, wl): return " ".join(reversed(wl)) |
| def longest_word(s, wl): return max(wl, key=len) |
| def shortest_word(s, wl): return min(wl, key=len) |
| def alphabetical_order(s, wl): return " ".join(sorted(wl, key=str.lower)) |
|
|
| def generate_split(pools, split_name, out_dir, per_lang=100): |
| tasks = [ |
| ("spell", spell, "word"), |
| ("reverse", reverse, "word"), |
| ("word_length", word_length, "word"), |
| ("first_letter", first_letter, "word"), |
| ("last_letter", last_letter, "word"), |
| ("vowel_count", vowel_count, "word"), |
| ("consonant_count", consonant_count, "word"), |
| ("remove_vowels", remove_vowels, "word"), |
| ("is_palindrome", is_palindrome, "word"), |
| ("word_count", word_count, "sentence"), |
| ("sentence_reverse", sentence_reverse, "sentence"), |
| ("longest_word", longest_word, "sentence"), |
| ("shortest_word", shortest_word, "sentence"), |
| ("alphabetical_order", alphabetical_order, "sentence"), |
| ] |
| all_items = [] |
| for name, fn, cat in tasks: |
| task_items = gen_task_base(name, fn, pools, cat, sentence_logic=(cat=="sentence"), per_lang=per_lang) |
| for i, item in enumerate(task_items): |
| item["id"] = f"{split_name}_{name}_{i:05d}" |
| all_items.append(item) |
| |
| with open(out_dir / f"{split_name}.jsonl", "w") as f: |
| for item in all_items: |
| f.write(json.dumps(item, ensure_ascii=False) + "\n") |
| return len(all_items) |
|
|
| def generate_all(): |
| random.seed(42) |
| validate_pools(WORD_POOLS) |
| out_dir = Path("data") |
| out_dir.mkdir(exist_ok=True) |
| |
| train_pools = [] |
| test_pools = [] |
| |
| for script, lang, words in WORD_POOLS: |
| tr, te = get_balanced_words(words) |
| train_pools.append((script, lang, tr)) |
| test_pools.append((script, lang, te)) |
| |
| |
| n_train = generate_split(train_pools, "train", out_dir, per_lang=50) |
| n_test = generate_split(test_pools, "test", out_dir, per_lang=50) |
| |
| print(f"Generated {n_train} train items and {n_test} test items.") |
|
|
| if __name__ == "__main__": |
| generate_all() |
|
|
