更新数据集发布状态和链接

README.md

@@ -6,23 +6,22 @@ colorTo: indigo
 sdk: static
 app_file: index.html
 pinned: false
-short_description: Archived SWE-agent RL experiment report
+short_description: Harbor 代码智能体强化学习实验档案
 ---
 
 # Harbor Qwen GRPO Experiment
 
-An evidence-backed archive of a tool-using code-agent reinforcement-learning project built with Harbor, rLLM, veRL, vLLM, GRPO and native critic/GAE PPO.
+这是一个以实验证据为基础的工具调用代码智能体强化学习项目档案。项目使用 Harbor、rLLM、veRL、vLLM、GRPO 和原生价值模型/GAE PPO，覆盖数据冻结、沙箱执行、训练、检查点评测和失败归因。
 
-The project completed its P0 evaluation on 2026-08-30. Under the unified 64K protocol, the original SFT model and the parent PPO checkpoint achieved the same aggregate results: 6/38 on the frozen SWE-Smith test panel and 9/27 on the SWE-bench Verified panel. The paired exact McNemar p-value was 1.0 on both panels, so this archive does not claim a verified overall PPO improvement.
+项目已于 2026 年 8 月 30 日完成 P0 评测。在统一 64K 协议下，原始 SFT 模型和父 PPO 检查点取得相同的总体结果：冻结 SWE-Smith 测试集均为 6/38，SWE-bench Verified 27 题均为 9/27。两个面板的任务级精确 McNemar 检验均为 `p=1.0`，因此本项目不宣称 PPO 带来了经过验证的整体能力提升。
 
-Open `index.html` through the Space application for the project overview, experiment evidence and links to the curated documentation.
+通过 Space 应用打开 `index.html`，可以查看项目概览、实验数据和整理后的项目文档。
 
-## Associated repositories
+## 关联仓库
 
-- Dataset: `keryszhan/harbor-swesmith-rl-artifacts` (publication pending)
-- Model: `keryszhan/klear-agentforge-8b-ppo-step20-adapter` (publication and license review pending)
+- 数据集：[keryszhan/harbor-swesmith-rl-artifacts](https://huggingface.co/datasets/keryszhan/harbor-swesmith-rl-artifacts)（已上传，当前为私有）
+- 模型：`keryszhan/klear-agentforge-8b-ppo-step20-adapter`（尚未发布，需先复核基础模型许可）
 
-## Publication scope
-
-This Space contains a curated documentation and source snapshot. It intentionally excludes credentials, runtime state, private registry material, upstream repository copies, container images and raw training checkpoints.
+## 发布范围
 
+本 Space 包含经过筛选的项目文档和源码快照，明确排除凭据、运行状态、私有镜像仓库材料、上游仓库副本、容器镜像和原始训练检查点。

index.html

@@ -136,7 +136,7 @@
         <div><p class="section-kicker">PUBLICATION BOUNDARY</p><h2>这是实验档案，不是效果宣称</h2></div>
         <ul>
           <li>不包含 `.env`、密钥、证书、Registry 密码或运行时状态。</li>
-          <li>最终 267 题数据由独立 Dataset 仓库发布；当前链接可能尚未开放。</li>
+          <li>最终 267 题数据已上传至独立的 <a href="https://huggingface.co/datasets/keryszhan/harbor-swesmith-rl-artifacts">Harbor SWE-Smith 强化学习数据集</a>，当前保持私有。</li>
           <li>PPO actor adapter 依赖 gated 基础模型，发布前需完成许可证复核。</li>
           <li>云资源已释放，文档中的部署命令属于可复现的历史运行手册。</li>
         </ul>
@@ -149,4 +149,3 @@
   </footer>
 </body>
 </html>
-