SakuraLLM provides specialized Japanese-to-Chinese translation models tailored for light novels and Galgame, targeting users who need high-quality, domain-specific translation. The models offer offline, self-deployable solutions with ACGN-style output, aiming to improve upon generic translation tools by understanding character relationships and universal attributes.

How It Works

SakuraLLM models are built upon open-source large language models, undergoing continued pre-training and fine-tuning with general Japanese corpora and domain-specific Chinese-Japanese data from light novels and Galgames. This approach leverages the strengths of base models while specializing them for nuanced translation, particularly in handling character pronouns and contextual understanding within the ACGN domain.

Quick Start & Requirements

• Installation: Models are available in GGUF format for use with compatible inference engines. Download links are provided on Hugging Face.
• Prerequisites: NVIDIA GPUs are recommended for optimal performance. VRAM requirements vary by model size, with 7B models needing ~8-10GB and 14B models requiring ~11-24GB. CPU inference is also supported.
• Resources: Setup involves downloading model weights and integrating with compatible inference tools like Sakura Launcher GUI or OneClickLLAMA.
• Documentation: Detailed tutorials for setup and usage are available in the repository Wiki and usage.md.

Highlighted Details

• Offers models ranging from ~2B to 32B parameters, based on Qwen and Qwen1.5 architectures.
• Supports a "GPT Dictionary" feature for consistent terminology and pronoun usage.
• Integrates with various translation tools like LunaTranslator, GalTransl, and manga-image-translator via a Sakura API.
• Recent updates (v1.0) focus on improved translation quality, accuracy, and handling of inline newlines.

Maintenance & Community

• Active development with regular model updates, including new versions based on Qwen2.5.
• A Telegram group is available for community discussion and support.
• The project acknowledges contributions from several individuals and projects.

Licensing & Compatibility

• Models are released under CC BY-NC-SA 4.0.
• Strictly prohibited for commercial use. All Sakura models and their derivatives are for learning and exchange purposes only.

Limitations & Caveats

• Commercial use is explicitly forbidden by the CC BY-NC-SA 4.0 license.
• Kaggle has reportedly banned SakuraLLM models, leading to permanent account bans for users.
• The project disclaims responsibility for any issues arising from the use of its models.