longtu-nllb-zh2ko

Fine-tuned facebook/nllb-200-distilled-600M for zh-CN → ko game localization translation.

Public model. License: cc-by-nc-4.0 (inherited from NLLB-200; non-commercial use only). Trained on a proprietary game localization corpus; game-domain terminology is encoded in the weights. The original corpus text is not distributed with this model.

Task

Sequence-to-sequence translation: Simplified Chinese (zho_Hans) → Korean (kor_Hang)

Base model

facebook/nllb-200-distilled-600M

Language pair

• Source: zh-CN (NLLB code: zho_Hans)
• Target: ko (NLLB code: kor_Hang)
• Direction: zh-CN → ko (unidirectional fine-tune)

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

repo = "SimpleJerry/longtu-nllb-zh2ko"
tag  = "earlystop-v1-ckpt48000"   # always pin to a specific published checkpoint

tokenizer = AutoTokenizer.from_pretrained(repo, revision=tag)
model     = AutoModelForSeq2SeqLM.from_pretrained(repo, revision=tag)

inputs = tokenizer("攻击力提升50%", return_tensors="pt", src_lang="zho_Hans")
output_ids = model.generate(
    **inputs,
    forced_bos_token_id=tokenizer.lang_code_to_id["kor_Hang"],
    num_beams=4,
    max_length=400,
)
print(tokenizer.decode(output_ids[0], skip_special_tokens=True))

Always pass revision=<tag> to pin to a specific published checkpoint. No token is required to pull this public repository.

Decoding defaults

See generation_config.json. Key parameters: num_beams=4, max_length=400.

Provenance

• corpus_sha256: 30D5C299828C10235AEE357E9333740913E55C291C5B07A45C0739E41818EA97
• split_seed: 42
• best_composite_metric (0.5·BLEU + 0.5·preservation_nospace): 0.6993
• Full training metadata: run_manifest.json (included in this repo)

Training governed by ADR-0020, ADR-0031. Distribution governed by ADR-0037.

Published tag

earlystop-v1-ckpt48000

---

Non-commercial use only (CC-BY-NC-4.0). Attribution required.