--- title: ChinaTravel canonical_url: "https://www.modelscope.cn/datasets/Cbphcr/ChinaTravel" md_url: "https://www.modelscope.cn/datasets/Cbphcr/ChinaTravel.md" repository: Cbphcr/ChinaTravel last_updated: 2026-09-15 license: cc-by-4.0 storage_size: "15 MB" downloads: 873 stars: 5 --- # ChinaTravel > ChinaTravel - Cbphcr 在 ModelScope 开源的数据集。ChinaTravel Query Dataset Cbphcr/ChinaTravel 是 ModelScope 魔搭社区上的数据集，存储大小 15 MB，采用 cc-by-4.0 许可。 - **Repository**: Cbphcr/ChinaTravel - **License**: cc-by-4.0 - **Storage size**: 15 MB - **Downloads**: 873 - **Stars**: 5 - **Last updated**: 2026-09-15 Source: https://www.modelscope.cn/datasets/Cbphcr/ChinaTravel --- # ChinaTravel Query Dataset This dataset is licensed under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). ChinaTravel is an open-ended travel-planning benchmark with compositional constraint validation for language agents. See the [paper](https://openreview.net/forum?id=0YRVlxY9BH), [Hugging Face paper page](https://huggingface.co/papers/2412.13682), [code](https://github.com/LAMDA-NeSy/ChinaTravel), and [bilingual sandbox database](https://huggingface.co/datasets/LAMDA-NeSy/ChinaTravel-Sandbox) ([ModelScope mirror](https://modelscope.cn/datasets/Cbphcr/ChinaTravel-Sandbox)) for the complete benchmark resources. ## Introduction For a given query, a language agent uses the sandbox tools to collect information and produces a travel plan in JSON format. A plan contains daily POIs, including restaurants, attractions, accommodations, and intercity transport, together with inner-city transport routes. The Phase 1 query records contain Chinese requests, English translations, and executable constraint programs. The Phase 2 config contains 2,000 English queries with their per-constraint English semantics and executable DSL. The static POI and transport data are distributed separately in the companion sandbox dataset linked above. ## Splits - **Default / Easy**: 300 queries with at most one extra constraint. - **Default / Medium**: 150 queries with complex constraints. - **Default / Human**: 154 human-authored queries. These are more diverse and may contain constraints not seen in the easy and medium splits. - **Default / Preference_base50**: 50 base queries used by the preference configuration. - **Test / Human1000**: 1,000 test queries. - **Preference / Preference0_base50**: Prefer more attractions. - **Preference / Preference1_base50**: Prefer less inner-city transport time. - **Preference / Preference2_base50**: Prefer less average transport time to restaurants. - **Preference / Preference3_base50**: Prefer more spending on food. - **Preference / Preference4_base50**: Prefer less spending on accommodation. - **Preference / Preference5_base50**: Prefer a shorter distance to a specified POI. - **TPC2026_phase2 / Full**: All 2,000 post-competition English queries covering the 53 published constraint templates. - **TPC2026_phase2 / Competition_test**: The 100-query subset used in the official Phase 2 competition evaluation. The same rows are identified in `full` by `official_phase2_evaluation=true`. Phase 2 requires the latest companion `ChinaTravel-Sandbox`; legacy sandbox snapshots are not compatible with the canonical English entity and concept labels. ## Record Layout - `uid`: unique query identifier. - `tag`: query category. - `start_city`: departure city. - `target_city`: destination city. - `days`: trip duration in days. - `people_number`: number of travelers. - `limit_rooms`: whether the query limits the number of rooms. - `limits_room_type`: whether the query limits room types. - `hard_logic_py`: executable Python constraint programs. - `nature_language`: Chinese natural-language request. - `nature_language_en`: English translation of the request. The keys below are only in preference config: - `preference`: Chinese preference description. - `preference_en`: English preference description. - `preference_py`: executable Python preference program. The Phase 2 config uses this public schema: - `uid`, `tag`, `start_city`, `target_city`, `days`, `people_number`; - `hard_logic_py`: ordered executable constraint programs; - `hard_logic_nl`: one English semantic statement per DSL program; - `nature_language`: the complete numbered English request; - `constraint_keys`: stable template identifiers aligned with both constraint lists; - `official_phase2_evaluation`: whether the record belongs to the 100-query competition evaluation subset. Because this repository's existing configs use the CSV builder, `hard_logic_py`, `hard_logic_nl`, and `constraint_keys` are JSON-encoded strings in the loaded Phase 2 table. Parse them with `json.loads`; the encoding preserves the ordered lists exactly. ```python import json from datasets import load_dataset phase2 = load_dataset("LAMDA-NeSy/ChinaTravel", "TPC2026_phase2", split="full") competition_test = load_dataset( "LAMDA-NeSy/ChinaTravel", "TPC2026_phase2", split="competition_test" ) record = phase2[0] hard_logic_py = json.loads(record["hard_logic_py"]) hard_logic_nl = json.loads(record["hard_logic_nl"]) constraint_keys = json.loads(record["constraint_keys"]) ``` Seed plans, source-plan paths, and generation traces are not included. See the [Phase 2 release notes](PHASE2_RELEASE_NOTES.md) and the machine-readable [audit report](phase2_audit_report.json) for validation and cleanup details. ## Data Maintenance On 2026-07-21, we applied two corrections reported in Hugging Face [Discussion #2](https://huggingface.co/datasets/LAMDA-NeSy/ChinaTravel/discussions/2) and [Discussion #3](https://huggingface.co/datasets/LAMDA-NeSy/ChinaTravel/discussions/3): - aligned the trip duration and English text of one `easy` Chengdu-Wuhan query; - replaced the unavailable Suzhou-Beijing airplane requirement in one `human` query with the available train requirement, consistently across its text and constraint program. The fixes are contained in Hub commit `802b18d`. No other query rows were changed by that cleanup. On 2026-08-18, we added the complete 2,000-query Phase 2 config after a post-competition audit. Natural-language wording was clarified without changing the meaning of the executable constraints, and logically redundant OR clauses were removed only when an identical branch was already independently required by the same query. Details and checksums are recorded in the linked release notes and audit report. On 2026-08-20, we corrected 28 Phase 1 records whose executable constraints were inconsistent with the query metadata or the companion sandbox inventory. The corrections cover canonical entity/concept labels, unsupported or contradictory categories, infeasible room-type and budget combinations, three undefined DSL variables, and one city mismatch. The canonical Shanghai cuisine label was also propagated to the same base query in all seven preference splits. No Phase 2 or `human1000` record required a corresponding change. All 604 corrected Phase 1 records have a plan that passes the current schema, commonsense, and hard-constraint evaluator. The exact affected UIDs, fields, file checksums, and full public-split audit are recorded in [`DATA_QUALITY_REPAIRS_2026_08.json`](DATA_QUALITY_REPAIRS_2026_08.json) and summarized in [`DATA_QUALITY_FIXES_2026_08.md`](DATA_QUALITY_FIXES_2026_08.md). Use these queries with release `2026.08.1` or later of the companion sandbox dataset. ## Citation If our paper or related resources prove valuable to your research, we kindly ask for citation. Please feel free to contact us with any inquiries. ```bibtex @inproceedings{shao2026chinatravel, title = {ChinaTravel: An Open-Ended Travel Planning Benchmark with Compositional Constraint Validation for Language Agents}, author = {Jie-Jing Shao and Bo-Wen Zhang and Xiao-Wen Yang and Baizhi Chen and Siyu Han and Pang Jinghao and Wen-Da Wei and Guohao Cai and Zhenhua Dong and Lan-Zhe Guo and Yu-Feng Li}, booktitle = {The Fourteenth International Conference on Learning Representations}, year = {2026}, url = {https://openreview.net/forum?id=0YRVlxY9BH} } ```