Model Card for LLlama-3.1-Taiwan-8B-Instruct-JPtoTaiwanLocal-4bit

This model is intended to be a customized fine-tuned version of yentinglin/Llama-3.1-Taiwan-8B-Instruct. The goal of this fine-tuning is to specialize in Japanese to Traditional Chinese (Taiwan-style) translation, especially for colloquial and adult content such as subtitles from Japanese AV media. The model will be adapted for local use with 4-bit quantization.

⚠️ This model card is submitted for access request approval only. Model fine-tuning has not yet been completed due to restricted access. This README will be updated accordingly once access is granted.

Model Details

Base model: yentinglin/Llama-3.1-Taiwan-8B-Instruct
Planned architecture: 8B instruction-tuned model with LoRA
Languages: Japanese (ja), Traditional Chinese (zh-TW)
License: Follows original Llama license
Planned training method: LoRA with bitsandbytes 4-bit quantization
Usage goal: Adult subtitle translation (JA → ZH-TW), emphasizing colloquial tone

Use Cases

Intended Use

Instruction-tuned bilingual assistant (Japanese → Taiwan Mandarin)
Translation of AV-style dialogue from Japanese to fluent Taiwanese-style Chinese
Subtitle generation or localization (e.g., fansubbing for adult content)
Internal chatbot or storywriting tools for 18+ applications

Out-of-Scope Use

Usage for minors or child-directed content
Use in non-consensual, unethical or illegal applications
Deployment in environments without content safeguards or disclaimers

Notes on Content

This model is planned to be trained on a set of manually curated Japanese adult subtitle translations into colloquial Taiwanese Mandarin. The content will be responsibly filtered to avoid non-consensual, underage, or illegal elements. The intended user base is strictly 18+.

Planned Training Setup

Training library: Hugging Face Transformers + PEFT + bitsandbytes
Quantization: load_in_4bit=True
Hardware: RTX 4080 16GB (local)
Data format: JSONL with instruction + input + output
Epochs: 3
LR: 2e-4

Repository

Model repository: https://huggingface.co/lipei0224/LLlama-3.1-Taiwan-8B-Instruct-JPtoTaiwanLocal-4bit

Contact

If you need more information or would like to collaborate, contact the author:
📧 [email protected]

This model card will be revised once full access is approved and the fine-tuning process is completed.