DataSpace

DataSpace is a benchmark for data agents that perform verifiable analytics over heterogeneous, task-local workspaces. Each task provides a natural-language question and a workspace containing a combination of CSV, JSON, SQLite, Markdown, PDF, and video artifacts. The required output is a complete tabular result.

DataSpace is also the official benchmark of the KDD Cup 2026 Data Agent Track.

Release policy

This release makes all 410 task inputs public. It includes gold.csv and the frozen evaluation configuration for a representative set of 60 tasks, allowing end-to-end local development and validation. References for the other 350 tasks are withheld for official full-benchmark evaluation.

The public-reference set was selected by deterministic constrained coverage optimization over source domains, question--workspace language combinations, evidence-carrier combinations, required modalities, analytical operations, answer shapes, numeric comparison modes, and ordering behavior. It covers all six modalities, all annotated analytical-operation families, and all evidence-carrier groups without repeated source samples.

Benchmark summary

Property	Value
Task inputs	410
Public reference packages	60
Context files	7,439
Uncompressed context size	15.0 GB
CSV files	2,809
JSON files	1,927
SQLite databases	402
Markdown documents	1,024
PDF documents	1,088
Videos	189

Chinese and English may occur across both the question and workspace artifacts. Tasks may require evidence discovery, structured extraction, joins across representations, filtering, aggregation, ranking, temporal reasoning, and document or video understanding.

Repository layout

.
├── README.md
├── CITATION.cff
├── LICENSE
├── data/
│   ├── tasks.jsonl
│   └── public_reference_tasks.json
└── release/
    ├── DataSpace-Benchmark.zip
    └── DataSpace-Benchmark.zip.sha256

data/tasks.jsonl is a lightweight index for the Hugging Face Dataset Viewer. It contains questions, workspace paths, available modality composition, file inventories, and a public-reference availability flag. It does not expose answer or solution-path metadata for withheld tasks.

Download

Install the Hugging Face Hub client and download the complete release:

pip install -U huggingface_hub
hf download HKUSTDial/DataSpace \
  release/DataSpace-Benchmark.zip \
  release/DataSpace-Benchmark.zip.sha256 \
  --repo-type dataset \
  --local-dir .

Verify and extract it:

cd release
sha256sum --check DataSpace-Benchmark.zip.sha256
unzip DataSpace-Benchmark.zip

For task task_N, the agent receives:

DataSpace-Benchmark/input/task_N/task.json
DataSpace-Benchmark/input/task_N/context/

and should write:

predictions/task_N/prediction.csv

Loading the task index

from datasets import load_dataset

tasks = load_dataset("HKUSTDial/DataSpace", split="test")
task = tasks[0]
print(task["question"])
print(task["workspace_path"])

Resolve workspace_path relative to the extracted DataSpace-Benchmark/ directory.

Evaluation

The official evaluator and baseline implementations are maintained in the DataSpace GitHub repository. Local evaluation is available for the 60 public-reference tasks. Official full-benchmark submissions are scored once over all 410 tasks, and missing predictions count as incorrect. Submission and leaderboard instructions will be published on the benchmark website.

Construction and quality control

DataSpace is constructed from EHRSQL and BULL Text-to-SQL source instances. The construction pipeline applies cross-language transformation, constraint-aware data sampling, modality routing and artifact rendering, and human review and data repair. Eleven domain experts cross-reviewed task solvability, question clarity, workspace evidence, reference answers, and evaluation configurations.

License

DataSpace is released under the MIT License. See LICENSE.

Citation

@misc{dataspace2027,
  title        = {DataSpace: Benchmarking Data Agents for Verifiable Analytics
                  over Heterogeneous Workspaces},
  author       = {Boyan Li and Zhuowen Liang and Yupeng Xie and Xiaotian Lin and
                  Tianqi Luo and Xinyu Liu and Yizhang Zhu and Zhangyang Peng and
                  Yuan Li and Zhengxuan Zhang and Jiayi Zhang and Nan Tang and
                  Guoliang Li and Yuyu Luo},
  year         = {2027},
  howpublished = {Hugging Face dataset},
  url          = {https://huggingface.co/datasets/HKUSTDial/DataSpace}
}