Spaces:
Running
Running
| title: Murasaki Project | |
| colorFrom: indigo | |
| colorTo: purple | |
| sdk: static | |
| emoji: 🔮 | |
| pinned: false | |
| app_file: README.md | |
| <div align="center"> | |
| <img src="https://github.com/soundstarrain/Murasaki-Translator/raw/main/GUI/resources/icon.png" width="160" height="160" style="border-radius: 50%; box-shadow: 0 4px 15px rgba(128, 0, 128, 0.3);"> | |
| <h1 style="font-size: 2.5em; margin-bottom: 10px;">🔮 Murasaki Project</h1> | |
| <p style="font-size: 1.2em; color: #6b7280;"> | |
| <b>System 2 Reasoning for Immersive ACGN Translation</b> | |
| </p> | |
| <p> | |
| 以 <b>System 2 推理范式</b>,重塑ACGN文学翻译体验 | |
| </p> | |
| <br> | |
| <!-- Badges --> | |
| <a href="https://github.com/soundstarrain/Murasaki-project" target="_blank"> | |
| <img src="https://img.shields.io/badge/GitHub-Murasaki_Project-181717?style=for-the-badge&logo=github" alt="GitHub"> | |
| </a> | |
| <a href="https://github.com/soundstarrain/Murasaki-Translator" target="_blank"> | |
| <img src="https://img.shields.io/badge/Tool-Murasaki_GUI-6B21A8?style=for-the-badge&logo=windows" alt="Translator Tool"> | |
| </a> | |
| <a href="https://github.com/soundstarrain/Murasaki-benchmark" target="_blank"> | |
| <img src="https://img.shields.io/badge/Benchmark-SOTA_Performance-blue?style=for-the-badge&logo=google-analytics" alt="Benchmark"> | |
| </a> | |
| </div> | |
| <br> | |
| <br> | |
| ## 🌌 Our Vision (愿景) | |
| **Murasaki Project** 致力于探索大语言模型在 **垂直领域文学翻译** 中的极限。 | |
| 翻译不仅仅是语言的转换,更是意境的重构。传统的直觉式(System 1)模型难以处理轻小说中复杂的长距离伏笔、细腻的人物语气和频繁的人称省略和转换。 | |
| 我们提出了 **System 2 Translation Paradigm** —— 通过引入显式思维链(**Chain-of-Thought**),让模型像人类资深译者一样: | |
| > **(先阅读语境 -> 分析语境 -> 落笔翻译)** | |
| 这种机制精准解决了 ACGN 翻译中"主语省略"、"人称混淆"及"风格漂移"的三大难题。 | |
| --- | |
| ## 🏆 Benchmark Performance | |
| 在 **Murasaki-ACGN Benchmark** 的综合评测中,**Murasaki-14B-v0.2** 取得了总分第一,**Murasaki-8B** 在8B的小参数下取得了总分第二和第三(均为4-bit量化版本评测得出) | |
| | Rank | Model | **Overall Avg** | Short | Long | | |
| | :--- | :--- | :--- | :--- | :--- | | |
| | 🥇 | **Murasaki-14B-v0.2** | **0.8545** | **0.8289** | **0.8801** | | |
| | 🥈 | Murasaki-8B-v0.1 | 0.8523 | 0.8269 | 0.8778 | | |
| | 🥉 | **Murasaki-8B-v0.2** | **0.8522** | **0.8271** | **0.8773** | | |
| | 4 | Gemini-3-Flash-Preview | 0.8512 | 0.8262 | 0.8765 | | |
| | 5 | Sakura-Qwen-2.5-14B | 0.8509 | 0.8282 | 0.8735 | | |
| --- | |
| ## 🧬 Model Matrix (模型矩阵) | |
| > **✨ Now Live:** 无需下载模型,点击 **[Online Demo](https://huggingface.co/spaces/Murasaki-Project/online-demo)** 在线体验模型。 | |
| | Model Name | Type | Size | VRAM (Est.) | Description | | |
| | :--- | :--- | :--- | :--- | :--- | | |
| | **[Murasaki-14B-v0.2](https://huggingface.co/Murasaki-Project/Murasaki-14B-v0.2)** | **BF16** | ~28GB | 32GB+ | **旗舰版**。全精度权重,最佳性能。 | | |
| | **[Murasaki-14B-v0.2-GGUF](https://huggingface.co/Murasaki-Project/Murasaki-14B-v0.2-GGUF)** | **GGUF** | 6.4~12GB | 12GB+ | **进阶本地版**。适合大显存用户。 | | |
| | **[Murasaki-8B-v0.2](https://huggingface.co/Murasaki-Project/Murasaki-8B-v0.2)** | **BF16** | ~16GB | 24GB+ | **标准版**。全精度权重,均衡之选。 | | |
| | **[Murasaki-8B-v0.2-GGUF](https://huggingface.co/Murasaki-Project/Murasaki-8B-v0.2-GGUF)** | **GGUF** | 3.6~6.3GB | 6GB+ | **推荐/轻量版**。兼容性最强,适合大多数显卡。 | | |
| ## 🛠️ Ecosystem (生态系统) | |
| 我们为模型提供了完整的工具链以降低使用门槛: | |
| * **[Murasaki Translator GUI](https://github.com/soundstarrain/Murasaki-Translator)** | |
| * 为Murasaki系列模型设计和优化的开箱即用的轻量级高性能翻译工具。 | |
| * **[Murasaki-ACGN Benchmark](https://github.com/soundstarrain/Murasaki-benchmark)** | |
| * 首个专注于长文本逻辑连贯性的ACGN翻译评测集。 | |
| --- | |
| ## 🤝 Credits & License | |
| Murasaki 系列模型基于 **SakuraLLM -(Qwen-3)-Base** 微调。 | |
| 我们严格遵守开源协议,所有模型权重均采用 **[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/)** 许可。 | |
| * Special thanks to the **SakuraLLM** team for their foundational work in the ACGN domain. | |
| * Powered by **Qwen** architecture. | |
| <br> | |
| <div align="center"> | |
| <p>Made with 💜 by the Murasaki Team</p> | |
| </div> |