from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model_name = "thefrigidliquidation/lightnovel-translate-Qwen2.5-32B"
model = AutoPeftModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

Prompt format

text

<|im_start|>system
Translate this text from Japanese to English.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Example:

text

<|im_start|>system
Translate this text from Japanese to English.<|im_end|>
<|im_start|>user
<GLOSSARY>
マイン : Myne
</GLOSSARY>
マイン、ルッツが迎えに来たよ<|im_end|>
<|im_start|>assistant
Myne, Lutz is here to take you home.

The glossary is optional. Remove it if not needed.

Text preprocessing

The Japanese text must be preprocessed with the following clean_string function that replaces some unicode characters with ASCII equivalents. Failure to do this may cause issues.

python

import ftfy

FTFY_ADDITIONAL_MAP = {
    "—": "--",
    "–": "-",
    "⸻": "----",
    "«": "\"",
    "»": "\"",
    "〝": "\"",
    "〟": "\"",
    "✧": "*",
    "✽": "*",
    "⬤": "*",
    "⭘": "*",
    "∴": "*",
    "∵": "*",
    "✩": "*",
    "【": "[",
    "】": "]",
    "「": "[",
    "」": "]",
    "〖": "[",
    "〗": "]",
    "〈": "<",
    "〉": ">",
    "《": "<<",
    "》": ">>",
}

def clean_string(text: str, strip: bool = True) -> str:
    config = ftfy.TextFixerConfig(normalization="NFC")
    s = ftfy.fix_text(text, config=config)
    s = "\n".join((x.strip() if strip else x.rstrip()) for x in s.splitlines())
    for b, g in FTFY_ADDITIONAL_MAP.items():
        s = s.replace(b, g)
    return s

Glossary

You can provide up to 30 custom translations for nouns and character names at runtime. Prefix your chapter with glossary terms (one per line) Japanese term : English term inside <GLOSSARY></GLOSSARY> tags.

For example, if you wish to have マイン translated as Myne you can construct the input prompt with:

python

glossary = [
    {"ja": "マイン", "en": "Myne"},
]
chapter_text = "マイン、ルッツが迎えに来たよ"

def make_glossary_str(glossary: list[dict[str, str]]) -> str:
    if glossart is None or len(glossary) == 0:
        return ""
    unique_glossary = {(term['ja'], term['en']) for term in glossary}
    terms = "\n".join([f"{ja} : {en}" for ja, en in unique_glossary])
    return f"<GLOSSARY>\n{terms}\n</GLOSSARY>\n"

user_prompt = f"{make_glossary_str(glossary)}{clean_string(chapter_text)}"

text

<GLOSSARY>
マイン : Myne
</GLOSSARY>
マイン、ルッツが迎えに来たよ