LocalSubs-EN-ZH-TW-0.6B

LocalSubsEN2TW-0.6B

A fine-tuned quantized model with domain specific tokenizer for cue-level, context-aware English → Taiwan Traditional Chinese subtitle translation, optimized for local real-time inference.

• Paper: Workload-Driven Optimization for On-Device Real-Time Subtitle Translation
• Repository: GitHub Code

Model Details

Intended Use

This model is designed for one task only:

Given recent English subtitle cues as context and the current English subtitle cue, output only the Taiwan Traditional Chinese translation of the current cue.

In scope:

• Real-time local subtitle overlay(avg. latency 52ms test on Macbook air M2)
• Cue-level streaming inference
• English → Taiwan Traditional Chinese (繁體中文・台灣用語)

Out of scope:

• General-purpose translation
• Long-document translation
• Simplified Chinese output
• General instruction following

Input / Output Format

CTX:
<0–3 previous English subtitle cues>

CUR:
<current English subtitle cue>

Output: Taiwan Traditional Chinese translation of CUR only.

Example:

CTX:
I sent you the file.
Check your inbox.

CUR:
You got it?

Expected output:

你收到了嗎？

Usage (llama.cpp)

./llama-cli \
  -m LocalSubs-EN-ZH-TW-0.6B-Q5_K_M.gguf \
  --temp 0.0 \
  -p "CTX:
I sent you the file.
Check your inbox.

CUR:
You got it?" \
  -n 64

Usage (Python · llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="SubtitleEN2TW-0.6B-Q5_K_M.gguf", n_ctx=256)

ctx_cues = ["I sent you the file.", "Check your inbox."]
cur_cue  = "You got it?"

ctx_block = "
".join(ctx_cues)
prompt = f"CTX:
{ctx_block}

CUR:
{cur_cue}"

out = llm(prompt, max_tokens=64, temperature=0.0, stop=["

"])
print(out["choices"][0]["text"].strip())

Training Data

The SFT dataset was built from two public subtitle corpora. The actual dataset is not redistributed due to upstream licensing constraints. Pipeline details and the diagnostic test set are available at Aiden1020/SubtitleEN2TW-SFT-Pipeline.

Filtering steps applied:

1. Text cleaning: remove HTML/ASS/VTT tags, music symbols, encoding-corruption markers (PUA chars, kana, Cyrillic), subtitle-group watermarks
2. Simplified Chinese rejection: OpenCC s2t diff ratio > 3% → dropped
3. CTX-leakage filter: discard samples where the Chinese target is disproportionately longer than the English cue, or closely matches the previous Chinese cue
4. English-echo filter: discard samples where the English source appears verbatim in the Chinese output
5. Length validation: English 1–200 chars, Chinese 1–120 chars
6. LLM-filtering: using Qwen3 35b as a judge to filter the split-cue.
7. Cue length balance: balancing the weight of long cue in the data distribution. Context size distribution in training set (0–3 cues): 30% / 30% / 25% / 15%.

A small manually curated diagnostic set (~3% mix-in) was included to cover leakage probes, short responses, ambiguous phrases, and Taiwan terminology.

Limitations

• Trained on subtitle-domain data only; may produce unnatural output for other domains.
• The training data is primarily from OpenSubtitles which contains crowd-sourced subtitles; quality varies.
• Output targets Taiwan Traditional Chinese style but may occasionally produce Hong Kong or neutral Traditional Chinese forms.
• Context crossing episode boundaries is treated as acceptable noise; no document-level segmentation is applied.
• Not suitable for general chat or instruction following.

Citation

If you use this model, please also cite the model:

@misc{wong2026localsubs,
  title         = {Workload-Driven Optimization for On-Device Real-Time Subtitle Translation},
  author        = {Tsz-To Wong},
  year          = {2026},
  eprint        = {2607.09957},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2607.09957},
  url           = {https://arxiv.org/abs/2607.09957}
}

also cite the base model:

@misc{luoyf2025lmt,
  title={NiuTrans.LMT: Toward Inclusive and Scalable Multilingual Machine Translation with LLMs},
  author={Yingfeng Luo, Ziqiang Xu, Yuxuan Ouyang, Murun Yang, Dingyang Lin, Kaiyan Chang, Tong Zheng, Bei Li, Peinan Feng, Quan Du, Tong Xiao, Jingbo Zhu},
  year={2025},
  eprint={2511.07003},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2511.07003}
}

License

This model is released under the Apache License 2.0.

See NOTICE for third-party attribution.