README.md

---
title: OpenSakura
emoji: 🌸
colorFrom: pink
colorTo: purple
sdk: static
pinned: false
---

<div align="center">
  <img src="https://raw.githubusercontent.com/OpenSakura/.github/main/profile/OpenSakura_512.png" width="200" alt="OpenSakura Logo">
  
  <h1>🌸 OpenSakura 🌸</h1>

  <p>
    <a href="https://github.com/OpenSakura"><img src="https://img.shields.io/badge/GitHub-OpenSakura-181717?logo=github&logoColor=white" alt="GitHub"></a>
    <a href="https://bench.opensakura.com/"><img src="https://img.shields.io/badge/📊_Benchmark-Dashboard-blue" alt="Benchmark Dashboard"></a>
    <a href="https://arena.opensakura.com/"><img src="https://img.shields.io/badge/⚔️_Arena-Translation-red" alt="Translation Arena"></a>
  </p>

  <p>
    <img src="https://readme-typing-svg.demolab.com?font=Fira+Code&weight=500&size=20&pause=1000&color=F194B4&center=true&vCenter=true&width=600&lines=Domain-specific+LLM+translation;Done+in+public%2C+with+receipts;Light+novels%2C+visual+novels%2C+and+more" alt="Typing SVG" />
  </p>
</div>

---

<h2 id="english">🌸 About Us</h2>

We build datasets, train models, run evaluations, and publish benchmarks for **literary and niche-domain translation**. Light novels, visual novels, galgames, web fiction — the domains where generic Machine Translation collapses into polite nonsense. 

At **OpenSakura**, we believe in open science and reproducible results. Everything we do is done in public, with receipts.

### 🎯 What We Do

- 📚 **Datasets** — Curated, schema-validated, with documented lineage and licensing. High-quality parallel corpora designed specifically for our target domains.
- 🔬 **Experiments** — Fully reproducible training runs with pinned revisions, deterministic setups, and meticulously logged metrics.
- 📊 **Benchmarks** — Comprehensive LLM-as-a-judge pairwise evaluations, establishing Elo-style rankings for translation models.
- ⚔️ **Arena** — Blind A/B human evaluation platform to gather high-fidelity preference data for RLHF and DPO.

---

### 🚀 Quick Links

| Resource | URL |
|---|---|
| 📊 **Benchmark Dashboard** | [bench.opensakura.com](https://bench.opensakura.com/) |
| ⚔️ **Translation Arena** | [arena.opensakura.com](https://arena.opensakura.com/) |
| 🐙 **GitHub** | [github.com/OpenSakura](https://github.com/OpenSakura) |

---

### 🤝 Get Involved

We welcome contributions of all kinds! Whether you want to contribute **datasets, tooling, evaluation scripts, documentation**, or just **file issues** when something breaks, we need your help!

👉 **Start here: [Contribution Guidelines](https://github.com/OpenSakura/Contributors-You-Dare-Not-Watch-This)**

*If you use OpenSakura artifacts and they help you, tell a friend. If they don't, tell us what failed (with examples).*

---

### 🌱 Origins

OpenSakura grew out of the [SakuraLLM](https://huggingface.co/SakuraLLM) community — the pioneering open-source project for Japanese-to-Chinese ACGN translation models. Members of that community came together to push the mission further: broader language pairs, rigorous reproducibility, public benchmarks, and a more open development process. We stand on their shoulders and are deeply grateful for the foundation they built.

---

<h2 id="中文介绍" align="center">🌸 中文介绍 🌸</h2>

<p align="center">
  <img src="https://readme-typing-svg.demolab.com?font=Fira+Code&weight=500&size=20&pause=1000&color=F194B4&center=true&vCenter=true&width=600&lines=%E4%B8%93%E6%B3%A8%E4%BA%8E%E7%89%B9%E5%AE%9A%E9%A2%86%E5%9F%9F%E7%9A%84+LLM+%E7%BF%BB%E8%AF%91;%E5%85%AC%E5%BC%80%E9%80%8F%E6%98%8E%EF%BC%8C%E6%9C%89%E8%BF%B9%E5%8F%AF%E5%BE%AA;%E8%BD%BB%E5%B0%8F%E8%AF%B4%E3%80%81%E8%A7%86%E8%A7%89%E5%B0%8F%E8%AF%B4%E5%8F%8A%E6%9B%B4%E5%A4%9A" alt="Typing SVG" />
</p>

OpenSakura 是一个专注于**特定领域大语言模型 (LLM) 翻译**的开源社区项目。我们致力于构建高质量数据集、训练专属模型、执行严格评测并发布权威基准——全面覆盖**轻小说、视觉小说 (Visual Novel)、Galgame、网络小说**等通用机器翻译极易“翻车”的垂直领域。

### 🎯 我们做什么

- 📚 **数据集** — 经过精心整理与严格校验，具备完整的来源追溯与明确的开源许可。
- 🔬 **实验** — 保证绝对可复现的训练流程，锁定数据集与模型版本，并记录完整的训练指标。
- 📊 **基准测试 (Benchmarks)** — 引入基于 LLM 作为裁判 (LLM-as-a-judge) 的成对评测机制，建立科学的 Elo 排名体系。
- ⚔️ **竞技场 (Arena)** — 开展盲评 A/B 人工评测，为后续的 RLHF 和 DPO 算法提供高质量的人类偏好数据。

---

### 🚀 快速链接

| 资源 | 链接 |
|---|---|
| 📊 **基准测试看板** | [bench.opensakura.com](https://bench.opensakura.com/) |
| ⚔️ **翻译竞技场** | [arena.opensakura.com](https://arena.opensakura.com/) |
| 🐙 **GitHub** | [github.com/OpenSakura](https://github.com/OpenSakura) |

---

### 🤝 参与贡献

我们热烈欢迎各种形式的开源贡献！无论是**提供数据集、开发工具链、编写评测脚本、完善文档**，还是仅仅**提交一个 issue** 报告 Bug，我们都需要你的力量！

👉 **从这里开始: [贡献指南](https://github.com/OpenSakura/Contributors-You-Dare-Not-Watch-This)**

*如果你觉得 OpenSakura 的产出对你有帮助，请推荐给你的朋友们！如果遇到糟糕的翻译结果，请带上具体的例子向我们反馈，帮助我们不断改进。*

---

### 🌱 社区起源

OpenSakura 脱胎于 [SakuraLLM](https://huggingface.co/SakuraLLM) 社区——那是日中 ACGN 翻译领域的开源先驱。怀揣着将这项事业推向更高峰的愿景，社区成员们再次集结：我们致力于**支持更丰富的语言对、贯彻更严格的可复现性标准、构建公开透明的基准评测，并推行更加开放的开发流程**。

我们站在前人的肩膀上，对 SakuraLLM 奠定的坚实基础致以最深的敬意与感谢。