tau2-zh-campus — the native Chinese campus domain for τ²-bench

The campus domain for τ²-bench / tau2-bench — a natively designed Chinese university academic-affairs domain (policy + simulated dual database + 50 graded tasks) for evaluating tool-using agents under the dual-control protocol. Scoring is pinned to the official tau2-bench v1.0.1 track.

Version v2.2.0 (2026-10-08): task-statement self-containment — the 15 tasks that cited an unpublished internal tool-specification document now carry inline guidance; authoring terms and evaluation jargon naturalized (68 string leaves total, itemized in the cleanup record shipped with this release); disclosure set landed in the domain README (policy-text eras, task-era detail, limitations, the M16 root-cause note, difficulty tiers, run-code provenance, cost caliber); student profile fields rectified (gender aligned to name semantics; 40 unique masked phone/ID values with disjoint tails — display fields only); manifest.json binds code to data (per-file sha256 + version twins); evidence governance: asset freeze policy, harness-metadata scrub (model reasoning traces kept with a redistribution note), the second-round calibration batch published, public data validator (validate-published-tasks.py, PASS 2063). 173 pytest. Scoring contract fields unchanged vs v2.1.0. Leaderboard numbers remain the v2.0.0-era as-run results (no re-runs). Data files updated: tasks.json, db.json (plus new SCHEMA.md, manifest.json). Code pin: tag campus-v2.2.0.

Version v2.1.0 (2026-10-07): task-statement cleanup — authoring markers removed from every string leaf of tasks.json (101 sites, including two user-scenario annotations under an explicit exception); M06's statement rewritten to its zero-write refusal semantics; policy v1.4.2 (a version footer added, article text unchanged); new consistency guards (full-leaf marker scan, task self-consistency lint, upstream-contract pins, solo-mode raise); 167 pytest. Scoring contract fields are byte-identical to v2.0.0: communicate_info, env_assertions, initial_state, per-task user_tools, and the core fields of every gold action (action_id/requestor/name/arguments). Four gold-action info text sites were cleaned (M05 ×1, H01 ×2, H15 ×1) — free-text annotations only, never read by the evaluator. Leaderboard numbers remain the v2.0.0-era as-run results (per-task statement deltas disclosed in the release notes). Data files updated: tasks.json, policy.md. Code pin: tag campus-v2.1.0.

Version v2.0.0 (2026-10-06): major data-contract release — internal QA metadata fully removed from the public files (description.notes dropped across all 50 tasks, issues keys deleted, P0x hint fragments scrubbed from agent-visible fields; embedded task copies inside the published raw runs normalized to the same data-contract (QA metadata stripped: notes/issues/P0x; assertion strings and trajectory bodies preserved)); policy.md meta-annotations scrubbed (article text unchanged) → v1.4.1; calibers recomputed — 54 env_assertions (25 tasks), 29 zero-write tasks (20 with explicit refusal semantics), 157 pytest. Raw runs & judge calibration published as GitHub Release assets. Code pin: tag campus-v2.0.0.

Version v1.1.5 (2026-10-06): release-engineering only — fork root-README.md banner (release tags are self-explanatory on clone) and runtime-readable tau2.domains.campus.__version__ / POLICY_VERSION / SCORING_PROTOCOL benchmark metadata. Dataset files unchanged since revision ed1e20e2…; code pin campus-v1.1.5.

Version v1.1.4 (2026-10-06): code-only follow-up — proxy authorization validity anchors to the signing time (art. 33); service-request diagnostics distinguish missing-ID vs missing-sign. Dataset files unchanged since revision ed1e20e2…; benchmark code pin is now tag campus-v1.1.4.

Version v1.1.3 (2026-10-06): internal QA issues metadata stripped from the public tasks.json (evaluation fields unchanged; gold-replay CI 50/50); proxy-pickup document closure and ticket target_grade_id slimming live in the code (tag campus-v1.1.3 — release & version manifest; the default branch tau2-zh is a showcase snapshot, not benchmark code).

Version v1.1.2 (2026-10-06): an undeclared hospital_level on medical uploads is returned-for-supplement instead of invalid (art. 12(2)); confirmation receipts agree with the settled state. Code-only; dataset files unchanged.

Version v1.1.1 (2026-10-05/06): consistency patch — illness deferrals now require two documents (diagnosis certificate + sick-leave note, policy art. 12(2)); special-channel submissions get a 10-working-day window (art. 9) with task H06 re-anchored to 2026-03-20; policy.md bumped to v1.4; enrollment-count bookkeeping is now authoritative-recomputed (reward-invariant recount on tasks M03/M10/M14/H12). Follow-up hardening (2026-10-06): pre-settle, user-side deadline guards, special-channel terminal transition, monotonic waitlist positions — zero live-impact on these dataset files. Full change log: PR #596, parts 1–5.

Version narrative: the notes above are an abridged summary — the single source of truth is the domain changelog in src/tau2/domains/campus/README.md (§8); the release notes, this dataset card, the project homepage, and the PR updates are aligned to it.

Provenance

• This dataset is the native-Chinese domain extension (campus) of sierra-research/tau2-bench (MIT) — not an official release; the source lives on the fork Zitrack/tau2-bench (dev/campus branch), filed upstream as PR #596.
• Policy, database, tasks, and personas are natively designed for the Chinese higher-education administration setting (not translated from an English domain); policy text is adapted and re-balanced from public university regulations.
• Domain package: src/tau2/domains/campus/ (data_model / user_data_model / tools / user_tools / environment); tests: tests/test_domains/test_campus/ (173 pytest, incl. the 50-task gold-replay CI and the deadline-guard matrix).

De-identification

• The university is fictional "Qingchuan University" (青川大学); the domain qcuniv.edu.cn is fictional and non-resolving (NXDOMAIN).
• All students, courses, tickets, grades, and records are fictional; no real names, student IDs, emails, school names, or personal data of any kind.
• Policy parameters are adapted from public regulations (MOE Order No. 41 and public university rules) and do not correspond to any real institution's current rules.
• Do not use this dataset against any real school or real individual.

Protocol

• Scoring is pinned to tau2-bench v1.0.1 (results are not comparable across versions); reward_basis = [DB, COMMUNICATE] + env_assertions, no RewardType.ACTION.
• Policy document v1.4.2 (v1.1–v1.4.2 increments; v1.4.1 = public-copy meta-annotation scrub and v1.4.2 = a version footer line, article text otherwise unchanged; earlier increments covered the workday-count alignment, the reschedule-confirmation wording, and a temporal-boundary clarification).
• 50 tasks (easy 15 / medium 20 / hard 15, incl. 29 zero-write tasks of which 20 carry explicit refusal semantics); the leaderboard protocol is 4 trials/task, seed=20261004, temperature=0, max_steps=60.

Files & fields

This repo is a τ²-bench domain asset bundle (policy + dual database + task definitions), not a row-oriented dataset — the Hugging Face Dataset Viewer intentionally does not render it; consume via the files below or the tau2-bench runtime.

File	Content	Key fields / structure
tasks.json	50 tasks	id (E/M/H numbering), description (task statement), user_scenario (persona script), initial_state.initialization_actions (e.g. bind_student), evaluation_criteria (reference actions / env_assertions / communicate_info entity strings / empty nl_assertions / reward_basis), user_tools (per-task student tool surface) — internal QA issues metadata stripped since v1.1.3
split_tasks.json	task splits	base (50) / easy (15) / medium (20) / hard (15), values are task-id lists
policy.md	the 36-article academic-affairs regulation (Qingchuan University)	time-bound trap clauses (add-drop window, deferral, maintenance window, month-end settlement, tiered appeals, …); the agent's single policy source
db.json	environment database	env (server-time anchor 2026-06-12) + 11 business tables: students / courses / course_offerings / enrollments / grades / exam_arrangements / deferral_requests / scholarships / scholarship_apps / certificates / tickets
user_db.json	student-side (app) data	3 tables: app_todos / pending_signatures / uploads — the action surface executed by the simulated user under dual-control

Grading

• DB: terminal-state hash comparison (the agent-side and user-side hashes must both match).
• COMMUNICATE: deterministic substring checks over the agent's replies (no whitespace / full-width normalization); every string in communicate_info is a must-convey entity.
• Both components gate the reward (product); no semantic / LLM judge in the reward path (see the calibration report — the 530/530 judge mechanism is fully rule-reproducible).

Disclosures

• Task-era disclosure: the anchor row was collected on the first generation of assertion strings (r1); the other four rows were collected after the assertion strings were revised. The embedded task copies inside each published run are preserved verbatim from run time apart from the v2.0.0 QA-metadata cleanup; they keep each row's as-run assertion strings and differ from the current published tasks.json at statement level (task-era disclosure) and are the primary evidence of each row's grading caliber.
• Zero-signal disclosure: 24 of the 50 tasks carry neither gold actions nor env_assertions — this is by design under the v1.0.1 contract (the DB component punishes spurious writes; process quality is not scored).
• The two headline validation figures must not be conflated: 530/530 = rule-reproducibility (the deterministic matcher reproduces every calibration decision), while 78.0% = human agreement on a constructively sampled calibration set — two different things.
• As-run policy disclosure: the policy text embedded in each run (per-sim) is the pre-v1.4 original; it has wording-level deviations from the implementation in force at the time, and grading semantics follow the implementation — which is the released semantics. Policy v1.4 was the text-alignment round that brought the text in line with the implementation. The run-level policy field has been normalized to the released version.

Citation

@misc{tau2-zh-campus,
  title        = {Tau2-ZH: A Native Chinese Campus Domain for $\tau^2$-bench},
  author       = {{Tau2-ZH Project}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ZitrackHF/tau2-zh-campus}},
  note         = {50 tasks; scoring pinned to tau2-bench v1.0.1; policy v1.4.2}
}

Please also cite the upstream τ²-bench: Si et al., τ²-bench: Evaluating Conversational Agents in a Dual-Control Environment, arXiv:2506.07982.

---

tau2-zh-campus — τ²-bench 中文高校教务域（campus）数据集

τ²-bench / tau2-bench 的 campus 域——原生设计的中文高校教务办事域（政策＋模拟双数据库＋50 道判分任务），用于在双控（dual-control）协议下评测工具调用型 Agent。计分钉死官方 tau2-bench v1.0.1 赛道。

版本 v2.2.0（2026-10-08）：任务陈述自足化——原引用未公开内部《工具规格》文档的 15 题改为内联指引；作者用语与评测语域自然化（共 68 个字符串叶，逐条见随本版发布的清洗记录）；披露集落域 README（政策三时代、任务时代明细、局限、M16 成因注记、难度分层、run-code 溯源、成本口径）；学生档案字段校正（性别对齐姓名语义；40 组唯一脱敏电话/证件号、尾号互斥——仅展示字段）；manifest.json 绑定代码与数据（逐件 sha256＋版本孪生）；证据治理：资产冻结纪律、harness 元数据洗刷（模型思维链保留并附再分发说明）、校准第二轮批次发布、公开数据校验器（validate-published-tasks.py，PASS 2063）。173 项 pytest。判分契约字段与 v2.1.0 相比零变化。榜单数字仍为 v2.0.0 时代 as-run 结果（未重跑）。数据文件更新：tasks.json、db.json（另新增 SCHEMA.md、manifest.json）。代码锚点：tag campus-v2.2.0。

版本 v2.1.0（2026-10-07）：任务陈述清洗——tasks.json 全部字符串叶清扫作者标记（101 处，含两处经明示例外的 user-scenario 标注）；M06 陈述重写为零写拒绝语义；政策 v1.4.2（新增版本脚注行，条款正文不变）；新增一致性守卫（全叶标记扫描、任务自洽 lint、上游契约钉、solo 抛错）；167 项 pytest。判分契约字段与 v2.0.0 逐字节一致：communicate_info、env_assertions、initial_state、逐题 user_tools，以及金标动作全部核心字段（action_id / requestor / name / arguments）；金标动作的 info 自由文本清理了 4 处（M05×1、H01×2、H15×1），evaluator 不读该文本。榜单数字仍为 v2.0.0 时代 as-run 结果（逐题陈述差异在 Release notes 披露）。数据文件更新：tasks.json、policy.md。代码锚点：tag campus-v2.1.0。

版本 v2.0.0（2026-10-06）：major 数据契约版本——公开文件内部 QA 元数据全量移除（50 题 description.notes 剥离、issues 键删除、agent 可见字段 P0x 提示片段清除；已发布原始跑批中的内嵌任务副本归一到同一数据契约——QA 元数据剥离：notes/issues/P0x；断言串与轨迹本体保留）；policy.md 元注记清除（条款正文零触碰）→ v1.4.1；口径重算——54 条 env_assertions（25 题）、29 道零写任务（20 道具明确拒绝语义）、157 项 pytest。原始跑批与判分校准以 GitHub Release assets 公开。代码锚点：tag campus-v2.0.0。

版本 v1.1.5（2026-10-06）：仅发布工程——fork 根 README.md 横幅（clone 发布 tag 即自解释）与运行时可读的 tau2.domains.campus.__version__ / POLICY_VERSION / SCORING_PROTOCOL 基准元数据。自修订 ed1e20e2… 起数据集文件零变化；代码锚点 campus-v1.1.5。

版本 v1.1.4（2026-10-06）：仅代码跟进——代领授权码有效期锚定签署时刻（第 33 条）；服务单据查询区分"缺证件"与"缺签署"。自修订 ed1e20e2… 起数据集文件零变化；基准代码锚点更新为 tag campus-v1.1.4。

版本 v1.1.3（2026-10-06）：公开 tasks.json 剥离内部 QA issues 元数据（评估字段不变；金标重放 CI 50/50）；代领证件闭环与工单 target_grade_id 瘦身随代码走（tag campus-v1.1.3——Release 与版本清单；默认分支 tau2-zh 为展示快照，非基准代码）。

版本 v1.1.2（2026-10-06）：医疗材料未申报医院等级改为"退回补报"而非"无效驳回"（第 12 条二）；确认回执与结算终态一致。仅代码变更，数据集文件零变化。

版本 v1.1.1（2026-10-05/06）：一致性修复——因病缓考须双材料（诊断证明＋病假条，政策第 12 条二）；特别通道提交获 10 个工作日窗（第 9 条），任务 H06 重锚至 2026-03-20；policy.md 升至 v1.4；选课计数改为权威重算（M03/M10/M14/H12 奖励不变重算）。后续加固（2026-10-06）：pre-settle、用户侧 deadline 守卫、特别通道终态迁移、候补位次单调——对本数据集文件零影响。完整变更日志：PR #596 part 1–5。

版本叙事：以上注记为摘要——唯一权威为域 README 变更记录（src/tau2/domains/campus/README.md §8）；Release notes、本数据卡、项目首页与 PR 更新与之对齐。

来源

• 本数据集是 sierra-research/tau2-bench（MIT 许可）的中文域扩展 campus，非官方发布；源码位于 fork Zitrack/tau2-bench（dev/campus 分支），并已经 PR #596 提交上游。
• 政策、数据库、任务、persona 均为中文高校教务场景原生设计（不是英文域翻译）；政策素材改写自公开高校规章制度并重新配比。
• 域代码五件套位于 src/tau2/domains/campus/（data_model / user_data_model / tools / user_tools / environment），测试位于 tests/test_domains/test_campus/（173 项 pytest，含 50 题全量金标重放 CI 与 deadline 守卫矩阵）。

虚构与脱敏声明

• 学校为虚构的"青川大学"（Qingchuan University）；域名 qcuniv.edu.cn 为虚构、不解析（NXDOMAIN）。
• 全部学生、课程、工单、成绩等记录均为虚构；无任何真实姓名、学号、邮箱、校名或个人信息。
• 政策参数为公开规章（教育部第 41 号令及公开高校制度）的改写重配比，不对应任何真实学校现行制度。
• 请勿将本数据集用于任何针对真实学校/真实个人的用途。

协议

• τ²-bench 计分协议钉 v1.0.1（跨版本结果不可比）；reward_basis = [DB, COMMUNICATE] + env_assertions，不使用 RewardType.ACTION。
• 政策文档 policy v1.4.2（v1.1–v1.4.2 增补；v1.4.1＝公开副本元注记清除、v1.4.2＝版本脚注行，条款正文其余零触碰；此前增补含工作日口径、缓考确认句对齐实现与时态边界澄清）。
• 任务规模：50 题（easy 15 / medium 20 / hard 15），其中 29 道零写任务（20 道具明确拒绝语义）；每题 4 trials、seed=20261004、temperature=0、max_steps=60 为项目榜单标准跑法。

文件说明

本仓库是 τ²-bench 域资产包（政策＋双数据库＋任务定义），非行式表格数据——Hugging Face Dataset Viewer 刻意不渲染；请经下方文件或 tau2-bench 运行时使用。

文件	内容	关键字段/结构
tasks.json	50 道任务	id（E/M/H＋编号）、description（题面）、user_scenario（persona 剧本）、initial_state.initialization_actions（如 bind_student）、evaluation_criteria（参考 actions / env_assertions / communicate_info 断言串 / nl_assertions 恒空 / reward_basis）、user_tools（按题裁剪的学生工具面）——v1.1.3 起内部 QA issues 元数据已剥离
split_tasks.json	任务分片	base（50）/ easy（15）/ medium（20）/ hard（15），值为 task id 数组
policy.md	青川大学教务办事规程（36 条）	时限坑条款（补退选窗口、缓考、维护窗口、月末结账、逐级申诉…）；agent 唯一政策来源
db.json	环境数据库	env（服务器时间锚 2026-06-12）＋ 11 张业务表：students / courses / course_offerings / enrollments / grades / exam_arrangements / deferral_requests / scholarships / scholarship_apps / certificates / tickets
user_db.json	学生侧（App）数据	3 张表：app_todos / pending_signatures / uploads —— dual-control 中由模拟用户执行的动作面

判分口径

• DB：终态校验和比对（agent 侧与 user 侧哈希必须同时一致）。
• COMMUNICATE：对 agent 全程回复做确定性子串匹配（空格/全半角不归一化），communicate_info 中每条串均为必现实体。
• 两项齐备方计 pass（乘积制）；判分不引入语义改写判定（详见项目校准报告，530/530 机理可由规则完全复现）。

披露

• 任务时代披露：锚点行采集于断言串第一代（r1），其余四行采集于换串后；已发布跑批中的内嵌任务副本除 v2.0.0 的 QA 元数据清理外自跑批时点逐字保留——保留各行 as-run 断言串，语句层与现行公布 tasks.json 不同（任务时代披露），即各行判分口径的原始凭证。
• 零信号披露：50 题中 24 题既无金标 actions 也无 env_assertions——系 v1.0.1 契约设计（DB 分量罚乱写，过程不评分）。
• 两个头条数字不可混读：530/530＝规则可复现性（确定性匹配器复现了全部校准判定），78.0%＝构造性抽样校准集上的人工一致率——两件事。
• as-run 政策披露：各次运行内嵌的政策文本（per-sim）为 v1.4 前原文，与当时实现存在措辞级偏差；判分语义以实现为准，即发布版语义。政策 v1.4 即把文本对齐到实现的那一轮。run-level 政策字段已归一为发布版。

引用格式

@misc{tau2-zh-campus,
  title        = {Tau2-ZH: A Native Chinese Campus Domain for $\tau^2$-bench},
  author       = {{Tau2-ZH Project}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/ZitrackHF/tau2-zh-campus}},
  note         = {50 tasks; scoring pinned to tau2-bench v1.0.1; policy v1.4.2}
}

同时请引用本体 τ²-bench：Si et al., τ²-bench: Evaluating Conversational Agents in a Dual-Control Environment, arXiv:2506.07982。