iflytek · Icecream0507 · Aug 26, 2026
diff --git a/.github/CODE_OF_CONDUCT.md b/.github/CODE_OF_CONDUCT.md
@@ -0,0 +1,42 @@
+# Code of Conduct
+
+Domux follows the iFLYTEK Open Source Community Code of Conduct.
+
+For the full text, please refer to our
+[Community Code of Conduct](https://github.com/iflytek/community/blob/master/code-of-conduct.md),
+which is adapted from the [Contributor Covenant](https://www.contributor-covenant.org), version 2.0.
+
+## Scope
+
+This Code of Conduct applies to all community spaces of this repository — issues, pull
+requests, discussions and code review — and applies when an individual is representing the
+project in public spaces.
+
+## Reporting
+
+To report abusive, harassing or otherwise unacceptable behaviour, please contact the
+iFLYTEK open source team at <ifly_opensource@iflytek.com>.
+
+All complaints will be reviewed and investigated promptly and fairly. Project maintainers
+are obligated to respect the privacy and security of the reporter of any incident.
+
+---
+
+# 行为准则
+
+Domux 遵循讯飞开源社区行为准则。
+
+完整内容请参阅
+[社区行为准则](https://github.com/iflytek/community/blob/master/code-of-conduct.md)，
+该准则改编自 [Contributor Covenant](https://www.contributor-covenant.org) 2.0 版本。
+
+## 适用范围
+
+本行为准则适用于本仓库的所有社区空间——包括 issue、pull request、讨论与代码评审，
+并适用于个人在公共场合代表本项目的场景。
+
+## 举报方式
+
+如需举报辱骂、骚扰或其他不可接受的行为，请联系讯飞开源团队：<ifly_opensource@iflytek.com>。
+
+所有举报都将被及时、公正地审查和调查。项目维护者有义务保护举报人的隐私与安全。
diff --git a/.github/ISSUE_TEMPLATE/bug_report.md b/.github/ISSUE_TEMPLATE/bug_report.md
@@ -0,0 +1,69 @@
+---
+name: 🐛 Bug report / 缺陷报告
+about: Report incorrect model output, a crash, or broken tooling / 报告模型输出错误、崩溃或工具链问题
+title: '[Bug] '
+labels: ['bug']
+---
+
+<!--
+Before filing, please search existing issues to avoid duplicates.
+提交前请先搜索已有 issue，避免重复。
+-->
+
+## What went wrong / 问题描述
+
+<!-- A clear description of the incorrect behaviour. / 清晰描述出现的错误行为。 -->
+
+## Category / 问题类别
+
+<!-- Keep the one that applies. / 保留适用的一项。 -->
+
+- [ ] Incorrect structured output (intent / device / room / action / value) / 结构化输出错误
+- [ ] Inference or deployment failure (vLLM / SGLang / Transformers) / 推理或部署失败
+- [ ] Training or reward plugin issue / 训练或奖励插件问题
+- [ ] Evaluation script (`eval/run_eval.py`) issue / 评测脚本问题
+- [ ] Documentation error / 文档错误
+- [ ] Other / 其他
+
+## Reproduction / 复现步骤
+
+**Input utterance / 输入指令:**
+
+```text
+<!-- e.g. 把客厅的灯调暗一点 -->
+```
+
+**Expected output / 期望输出:**
+
+```json
+
+```
+
+**Actual output / 实际输出:**
+
+```json
+
+```
+
+<!--
+For non-output bugs, give the exact commands and the full error traceback instead.
+若不是输出类问题，请改为提供完整命令与报错栈。
+-->
+
+## Environment / 运行环境
+
+| Item / 项目 | Value / 值 |
+| --- | --- |
+| Model source / 模型来源 | <!-- Hugging Face / ModelScope / local fine-tune --> |
+| Model version / 模型版本 | <!-- e.g. Domux-Gemma-4-E2B-it --> |
+| Inference backend / 推理后端 | <!-- vLLM 0.22.0 / SGLang 0.5.12 / Transformers --> |
+| Python / OS | |
+| GPU / hardware / 硬件 | |
+
+## Additional context / 补充信息
+
+<!--
+Logs, screenshots, sampling parameters (temperature, top_p), or a link to a case
+under `cases/` that demonstrates the problem.
+日志、截图、采样参数（temperature、top_p），或 `cases/` 下可复现该问题的案例链接。
+-->
diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml
@@ -0,0 +1,17 @@
+blank_issues_enabled: true
+contact_links:
+  - name: 📖 Documentation / 项目文档
+    url: https://github.com/iflytek/domux#readme
+    about: Quick start, benchmark, deployment and output specification. / 快速开始、评测、部署与输出规范。
+  - name: 🤗 Model on Hugging Face / Hugging Face 模型
+    url: https://huggingface.co/iFlytekOpenSource/Domux
+    about: Model card, weights and usage notes. / 模型卡、权重与使用说明。
+  - name: 🔧 Model on ModelScope / ModelScope 模型
+    url: https://modelscope.cn/models/iflytek/domux
+    about: Mirror for users in mainland China. / 面向中国大陆用户的镜像。
+  - name: 🔒 Report a security vulnerability / 报告安全漏洞
+    url: https://github.com/iflytek/community/blob/master/SECURITY.md
+    about: Please do NOT open a public issue for security problems. / 安全问题请勿公开提交 issue。
+  - name: 💬 iFLYTEK Open Source Community / 讯飞开源社区
+    url: https://github.com/iflytek/community
+    about: Contribution guide and code of conduct. / 贡献指南与行为准则。
diff --git a/.github/ISSUE_TEMPLATE/feature_request.md b/.github/ISSUE_TEMPLATE/feature_request.md
@@ -0,0 +1,57 @@
+---
+name: ✨ Feature request / 功能建议
+about: Suggest a new capability, device type, or tooling improvement / 建议新能力、新设备类型或工具链改进
+title: '[Feature] '
+labels: ['enhancement']
+---
+
+<!--
+For HER Hack-Astron challenge submissions, please use the dedicated template instead.
+如需提交 HER Hack-Astron 赛题，请改用对应模板。
+-->
+
+## Problem / 要解决的问题
+
+<!--
+What is missing or hard to do today? Describe the user-facing scenario rather than the
+solution. 当前缺少什么，或者什么事情很难做到？请先描述用户场景，而不是解决方案。
+-->
+
+## Proposed solution / 建议方案
+
+<!-- What you would like to see. / 你希望看到的实现。 -->
+
+## Area / 涉及范围
+
+- [ ] Control capability — new device type, attribute, or action / 控制能力：新设备类型、属性或动作
+- [ ] Semantic understanding — dialect, accessibility, context, negation / 语义理解：方言、无障碍、上下文、否定
+- [ ] Model / training — data, reward, quantization, on-device latency / 模型与训练：数据、奖励、量化、端侧时延
+- [ ] Inference / deployment — backend or serving support / 推理与部署：后端或服务支持
+- [ ] Evaluation / dataset / 评测与数据集
+- [ ] Documentation or developer experience / 文档或开发体验
+- [ ] Other / 其他
+
+## Example / 示例
+
+**Input utterance / 输入指令:**
+
+```text
+
+```
+
+**Desired structured output / 期望的结构化输出:**
+
+```json
+
+```
+
+## Alternatives considered / 已考虑的替代方案
+
+<!-- Optional. / 选填。 -->
+
+## Additional context / 补充信息
+
+<!--
+Related issues, papers, datasets, or upstream references.
+相关 issue、论文、数据集或上游参考资料。
+-->
diff --git a/.github/ISSUE_TEMPLATE/question.md b/.github/ISSUE_TEMPLATE/question.md
@@ -0,0 +1,35 @@
+---
+name: ❓ Question / 使用咨询
+about: Ask about usage, deployment, training, or evaluation / 咨询使用、部署、训练或评测相关问题
+title: '[Question] '
+labels: ['question']
+---
+
+<!--
+Please check these first — they answer most questions:
+提问前请先查阅以下资料，多数问题已有答案：
+
+- README.md / README_zh.md
+- docs/output-spec.md / docs/output-spec.zh.md  (output contract / 输出规范)
+- eval/DATASET_README.md  (dataset & evaluation / 数据集与评测)
+- cases/README.md  (case studies / 案例)
+-->
+
+## Question / 问题
+
+<!-- What would you like to know? / 你想了解什么？ -->
+
+## What I have tried / 已经尝试过的
+
+<!--
+Commands you ran, docs you read, and what happened.
+执行过的命令、查阅过的文档，以及实际结果。
+-->
+
+## Environment / 运行环境 (if relevant / 如相关)
+
+| Item / 项目 | Value / 值 |
+| --- | --- |
+| Model source / 模型来源 | |
+| Inference backend / 推理后端 | |
+| Python / OS | |
diff --git a/THIRD_PARTY_LICENSES.md b/THIRD_PARTY_LICENSES.md
@@ -0,0 +1,109 @@
+# Third-Party Licenses
+
+This document inventories the third-party components that Domux depends on, together with a
+license risk assessment.
+
+It complements [`NOTICE`](NOTICE), which carries the attribution and pass-through obligations
+required for redistribution. Where the two overlap (notably the Gemma model weights), `NOTICE`
+is the authoritative statement.
+
+- **Domux's own code** (training scripts, reward plugins, evaluation tooling) is licensed under
+  the Apache License, Version 2.0 — see [`LICENSE`](LICENSE).
+- **Domux's model weights** are a Gemma derivative and are **not** covered by Apache-2.0.
+  See [Model weights](#model-weights) below.
+
+**Scope of this inventory.** It covers the *direct*, first-order dependencies declared in this
+repository. It does not enumerate transitive dependencies. Licenses were verified against
+package metadata published on PyPI on **2026-08-26**; declared metadata can change between
+releases, so re-verify when bumping a pin.
+
+---
+
+## Training
+
+Declared in [`training/requirements.txt`](training/requirements.txt).
+
+| Component | Constraint | License (as declared) | Project |
+| --- | --- | --- | --- |
+| `ms-swift[llm]` | `>=3.0.0` | Apache-2.0 | [modelscope/ms-swift](https://github.com/modelscope/ms-swift) |
+| `torch` | `>=2.0.0` | `Apache-2.0 AND Apache-2.0 WITH LLVM-exception AND BSD-2-Clause AND BSD-3-Clause AND BSL-1.0 AND MIT` | [pytorch/pytorch](https://github.com/pytorch/pytorch) |
+| `transformers` | `>=4.40.0` | Apache-2.0 | [huggingface/transformers](https://github.com/huggingface/transformers) |
+| `peft` | `>=0.10.0` | Apache-2.0 | [huggingface/peft](https://github.com/huggingface/peft) |
+
+> `torch` declares a composite SPDX expression because the distributed wheels bundle
+> third-party components under several licenses. The PyTorch project's own source license is
+> BSD-3-Clause. Anyone **redistributing** PyTorch binaries (rather than installing them from
+> PyPI) should also ship PyTorch's bundled third-party notices.
+
+## Inference and deployment
+
+Declared in the Quick Start / Deployment sections of [`README.md`](README.md).
+
+| Component | Constraint | License (as declared) | Project |
+| --- | --- | --- | --- |
+| `vllm` | `==0.22.0` | Apache-2.0 | [vllm-project/vllm](https://github.com/vllm-project/vllm) |
+| `sglang[all]` | `==0.5.12` | Apache-2.0 | [sgl-project/sglang](https://github.com/sgl-project/sglang) |
+
+> These are alternative backends — a deployment normally installs one, not both. The `[all]`
+> extra of `sglang` pulls in a substantially larger dependency set than the base package; audit
+> it separately if you redistribute a bundled environment.
+
+## Evaluation and tooling
+
+| Component | Used by | License (as declared) | Project |
+| --- | --- | --- | --- |
+| `requests` | [`eval/run_eval.py`](eval/run_eval.py) | Apache-2.0 | [psf/requests](https://github.com/psf/requests) |
+| `modelscope` | model download ([`README.md`](README.md)) | Apache-2.0 | [modelscope/modelscope](https://github.com/modelscope/modelscope) |
+
+[`scripts/validate_cases.py`](scripts/validate_cases.py) uses only the Python standard library
+and introduces no third-party dependency.
+
+---
+
+## Model weights
+
+Domux (`Domux-Gemma-4-E2B-it`) is a fine-tuned derivative of Google's Gemma model.
+
+| Component | Terms |
+| --- | --- |
+| Gemma / Domux model weights | [Gemma Terms of Use](https://ai.google.dev/gemma/terms) + [Gemma Prohibited Use Policy](https://ai.google.dev/gemma/prohibited_use_policy) |
+
+The Gemma Terms of Use are **not** an OSI-approved open source license. They impose use
+restrictions that must be passed through to downstream recipients. The full obligations —
+providing the terms, displaying a "built from Gemma" notice, and passing through the
+restrictions — are set out in [`NOTICE`](NOTICE).
+
+"Gemma" is a trademark of Google LLC.
+
+---
+
+## Risk assessment
+
+| # | Area | Risk | Assessment |
+| --- | --- | --- | --- |
+| 1 | Direct Python dependencies | **Low** | All are permissive (Apache-2.0, BSD, MIT). No copyleft (GPL/LGPL/AGPL) or source-available/non-commercial license appears in the direct dependency set. All are compatible with distributing this repository's code under Apache-2.0. |
+| 2 | Model weights | **Medium — the principal obligation** | The weights are governed by the Gemma Terms of Use, not Apache-2.0. Redistribution requires passing through the terms and the Prohibited Use Policy. Users who assume "Apache-2.0 repository ⇒ Apache-2.0 weights" will be wrong; `NOTICE` and the model cards state the correct terms. |
+| 3 | Transitive dependencies | **Unassessed** | Not enumerated here. `torch`, `vllm` and `sglang[all]` each pull in large trees. Anyone shipping a bundled environment or container image should run a full transitive scan rather than relying on this document. |
+| 4 | Bundled binaries in `torch` | **Low** | Handled upstream by PyTorch's own notices; relevant only when redistributing PyTorch binaries. |
+| 5 | Version drift | **Low, ongoing** | Training dependencies use `>=` constraints, so the resolved version — and in principle its declared license — can change over time. Re-verify when pinning for a release. |
+| 6 | Datasets | **See dataset docs** | Evaluation data ships in [`eval/`](eval); its provenance and terms are described in [`eval/DATASET_README.md`](eval/DATASET_README.md). Contributed case studies must satisfy the data-source requirements in [`cases/README.md`](cases/README.md). |
+
+**Summary.** The code-side license position is clean: every direct dependency is permissive and
+Apache-2.0-compatible. The meaningful compliance obligation lies with the **model weights**,
+which are Gemma-derived and carry use restrictions that must travel with any redistribution.
+
+---
+
+## Maintaining this document
+
+When adding, removing, or re-pinning a dependency, update the tables above.
+
+To regenerate a full inventory including transitive dependencies:
+
+```bash
+pip install pip-licenses
+pip-licenses --format=markdown --with-urls --with-license-file
+```
+
+Verify anything the tooling reports as `UNKNOWN` against the project's own `LICENSE` file
+rather than recording it as unknown.