# 云小鹤本地运行时 r7 / YunXiaoHe local runtime r7

版本 / Version: `20261003-r7` · [开始使用 / Get started](getting-started.zh-CN.md) · [English guide](getting-started.md)

## 中文

用你熟悉的 Codex、DSH 或 Worker 描述任务，云小鹤在本机完成数据检查、统计、RISA 建模和预测。YH 账号负责授权；你的 Harness 继续使用自己的模型和额度，模型 key 不经过 YH。

### 本次更新

- 修复 Windows 长路径安装，不需要修改系统注册表；修复 Linux 归档与安装器的路径兼容问题。
- 增加原始 TabArena 二分类协议模式，保留原始特征类型、类别顺序、五折划分、选模指标和 T1 规则。
- 修复交付模型再次训练和预测时的类别顺序及 CSV 文本值，保留 `01`、`1`、`1.0`、`NA` 和行标识。
- 普通建模继续默认精度优先。分数相同时，按原算法优先选择五折平均模型包更小的候选，再按固定候选 ID 排序。T1 的 `on` 与 `auto` 分别处理。

### 验证结果

原始 TabArena 的 **30 个二分类任务全部复现**：选中的模型、五折行划分、编码器和最终预测排名一致，涵盖 **472 个候选**，包括 112 个被 T1 条件拒绝的候选。

原记录的宏平均 ROC-AUC 为 **0.816690**。本次通过相同测试样本上的预测排名验证成绩一致，没有重新读取测试标签。28 项原始分数逐字节相同；另两项 T1 任务只有最大约 `2.05e-14` 的浮点差异，排名不变。

完整复现使用原 DGX ARM64 环境。Windows x64 与 Linux x86_64 分别检查安装、原生计算、模型保存与复用。训练中的 OOF 指标用于选模，不是独立测试成绩；本次没有新增 token 或费用对比。

### 使用与平台

安装器根据官网当前发布清单选择版本。已有用户更新 Skill 后，重新运行：

```text
python scripts/yxh.py runtime install
python scripts/yxh.py runtime doctor
```

Windows 可以使用 `py -3`，Linux 使用 `python3`。YH 账号需要云小鹤和本地应用权限。Linux 需要 Ubuntu 24.04 或兼容的 glibc 2.39+ 环境；Windows 包为未签名的便携运行时。请通过官方渠道下载并校验摘要。macOS/ARM64 没有本次安装包。

平常只需说明目标列、哪些字段在预测时可用，以及数据是否有时间或分组关系。只有明确复现原始协议时，才使用 [TabArena 模式](../yunxiaohe-client/references/local-data.md#original-tabarena-protocol)。不必把普通任务改成基准测试。

## English

Describe a task in Codex, DSH or your own Worker. YunXiaoHe handles data checks, statistics, RISA modeling and prediction locally. Your YH account authorizes the product; your Harness keeps its selected model, provider key and usage account.

### What's new

- Windows long-path installation works without registry changes; Linux archives use paths compatible with the installer.
- An explicit original TabArena binary profile preserves feature types, class order, folds, selection metrics and T1 behavior.
- Delivered models retain class order and CSV tokens such as `01`, `1`, `1.0` and `NA` across retraining and prediction.
- Ordinary modeling remains accuracy-first. Score ties follow the original mean-fold packet size and fixed candidate-ID ordering. Explicit T1 `on` and conditional `auto` are distinct.

### Validation

All **30 original binary tasks** have matching selected models, folds, encoders and final prediction rankings across **472 candidates**, including 112 rejected T1 candidates. The recorded macro ROC-AUC is **0.816690**. Matching rankings on the same test instances establish score parity without rereading test labels. Raw scores are byte-identical for 28 tasks; the two T1 tasks differ by at most about `2.05e-14`, with unchanged rankings.

The full replay used the original DGX ARM64 environment. Windows/Linux x64 have separate installation, native-computation and model-reuse tests. Training OOF metrics support selection, not independent testing. This release does not add a new token or cost comparison.

### Getting started

Update the Skill, then run the two commands above to install the version currently published in the official catalog. Windows can use `py -3`; Linux can use `python3`. Your account needs YunXiaoHe and local-app access. Linux requires Ubuntu 24.04 or compatible glibc 2.39+. The Windows portable executable is unsigned; use official downloads and verify their digests. No macOS/ARM64 package is provided.

For ordinary work, describe the target, prediction-time features and any time/group relationships. Use the [original-protocol profile](../yunxiaohe-client/references/local-data.md#original-tabarena-protocol) only when that is your intended experiment.

Support: [bot@yh-intel.com](mailto:bot@yh-intel.com)