| """ |
| core/lang.py — Language detection and chunk translation utilities. |
| |
| If a source chunk is not in English, the LLM struggles to follow |
| "write in English only" instructions because the French/Spanish tokens |
| in the context dominate the probability distribution at each decoding step. |
| Pre-translating the chunk removes this gravitational pull entirely. |
| """ |
|
|
| import re |
| from functools import lru_cache |
|
|
| |
| _FR_INDICATORS = frozenset([ |
| "le", "la", "les", "de", "du", "des", "un", "une", "est", "en", "et", |
| "avec", "dans", "sur", "pour", "que", "qui", "se", "au", "aux", "par", |
| "ou", "ne", "pas", "plus", "son", "sa", "ses", "leur", "leurs", "lui", |
| "ils", "elles", "nous", "vous", "je", "tu", "il", "elle", "ce", "cet", |
| "cette", "ces", "mon", "ma", "ta", "sont", "ont", "une", "comme", "aussi", |
| "mais", "donc", "car", "si", "tout", "tous", "toute", "toutes", "quel", |
| "quelle", "quels", "quelles", "dont", "très", "aussi", "puis", |
| ]) |
|
|
|
|
| def is_english(text: str) -> bool: |
| """Return True if *text* appears to be in English.""" |
| words = set(re.findall(r'\b[a-zA-ZÀ-ÿ]{2,}\b', text.lower())) |
| french_hits = len(words & _FR_INDICATORS) |
| |
| return french_hits < 5 |
|
|
|
|
| @lru_cache(maxsize=64) |
| def _cached_translate(chunk: str) -> str: |
| from model.llm import get_llm |
| llm = get_llm() |
| prompt = ( |
| "Translate the following text to English. " |
| "Output ONLY the English translation, nothing else.\n\n" |
| "Text:\n" + chunk + "\n\nTranslation:" |
| ) |
| max_tok = min(600, max(80, len(chunk.split()) * 2)) |
| result = llm.generate(prompt, max_new_tokens=max_tok, temperature=0.1).strip() |
| |
| if len(result) < 20: |
| return chunk |
| return result |
|
|
|
|
| def ensure_english(chunk: str) -> str: |
| """Return an English version of *chunk*, translating only if necessary.""" |
| if is_english(chunk): |
| return chunk |
| return _cached_translate(chunk) |
|
|