SakuraLLM/SakuraLLM

What it solves

SakuraLLM provides a specialized set of open-source, self-deployable translation models designed specifically for the ACGN (Anime, Comic, Game, Novel) domain. It addresses the challenge of translating Japanese light novels and Galgames (visual novels) into Simplified Chinese, focusing on maintaining the specific stylistic nuances, correct pronoun usage, and the preservation of technical control characters (like \n) often found in game scripts.

How it works

The project utilizes a series of open-source base models (primarily from the Qwen series, including Qwen 1.5 and Qwen 2.5) and applies continual pre-training and Supervised Fine-Tuning (SFT) and RLHF. The models are optimized for Japanese-to-Chinese translation with a focus on "Universal Character and Relationship Attributes" to improve the accuracy of pronouns and context understanding.

Who it’s for

• Translators and Fans: Users who want high-quality, ACGN-style translations of Japanese visual novels, light novels, and manga.
• Developers: Those building translation tools (e.g., LunaTranslator, GalTransl) that require a specialized backend for Japanese-to-Chinese translation.
• Self-hosters: Users who want to run translation models locally and offline for privacy or control.

Highlights

• Domain-Specific Optimization: Specifically tuned for light novels and Galgames, ensuring the style is appropriate for the genre.
• Glossary Support: Supports "GPT Dictionaries" (terminology lists) to ensure consistency in proper nouns and character names.
• Multiple Model Sizes: Offers a range of model sizes (from 1.5B to 32B parameters) to accommodate different hardware capabilities.
• Control Character Preservation: High ability to retain line breaks and other control symbols essential for game script translation.
• OpenAI API Compatibility: The API backend supports OpenAI-formatted requests, making it easy to integrate into existing tools.