| | --- |
| | language: |
| | - zh |
| | license: apache-2.0 |
| | task_categories: |
| | - question-answering |
| | - tabular-classification |
| | - text-generation |
| | tags: |
| | - data-analytics |
| | - agents |
| | - document-understanding |
| | - benchmark |
| | pretty_name: AIDABench |
| | --- |
| | |
| | # Dataset Card for AIDABench |
| |
|
| | ## Dataset Summary |
| |
|
| | **AIDABench** is a benchmark for evaluating AI systems on **end-to-end data analytics over real-world documents**. It contains **600+** diverse analytical tasks grounded in realistic scenarios and spans heterogeneous data sources such as **spreadsheets, databases, financial reports, and operational records**. Tasks are designed to be challenging, often requiring multi-step reasoning and tool use to complete reliably. |
| |
|
| |  |
| |
|
| | *Figure 1: Overview of the AIDABench evaluation framework.* |
| |
|
| | ## Supported Tasks and Evaluation Targets |
| |
|
| | AIDABench focuses on practical document analytics workflows where a model/agent must read files, reason over structured data, and produce a final deliverable. |
| |
|
| | ### Task Categories |
| |
|
| | The dataset is organized around three primary capability dimensions: |
| |
|
| | - **File Generation (43.3%)** |
| | Data wrangling and transformation tasks such as filtering, normalization, deduplication, joins, and cross-sheet linkage, with outputs as generated files (e.g., spreadsheets). |
| |
|
| | - **Question Answering (QA) (37.5%)** |
| | Analytical queries such as aggregation, averages, ranking, comparisons, and trend analysis, with outputs as final answers. |
| |
|
| | - **Data Visualization (19.2%)** |
| | Chart creation/adaptation tasks (e.g., bar/line/pie) including style requirements and presentation constraints, with outputs as figures or chart files. |
| |
|
| |  |
| |
|
| | *Figure 2: Example evaluation scenarios for QA, Data Visualization, and File Generation.* |
| |
|
| | ### Task Complexity |
| |
|
| | Tasks are stratified by the number of expert-level reasoning steps required: |
| |
|
| | - **Easy (29.5%)**: ≤ 6 steps |
| | - **Medium (49.4%)**: 7–12 steps |
| | - **Hard (21.1%)**: ≥ 13 steps |
| | - **Cross-file Reasoning**: 27.4% of tasks require reasoning over multiple input files (up to 14 files). |
| |
|
| | ### Data Formats |
| |
|
| | Most inputs are tabular files (xlsx/csv dominate), complemented by **DOCX** and **PDF** formats to support mixed-type document processing. |
| |
|
| | ## Evaluation Framework |
| |
|
| | All models are evaluated under a unified **tool-augmented protocol**: the model receives task instructions and associated files, and can execute **arbitrary Python code** within a **sandboxed environment** to complete the task. |
| |
|
| | To align with task categories, AIDABench uses three dedicated **LLM-based evaluators**: |
| |
|
| | 1. **QA Evaluator** |
| | A binary judge that determines whether the produced answer matches the reference (under the benchmark’s scoring rules). |
| |
|
| | 2. **Visualization Evaluator** |
| | Scores both **correctness** and **readability** of generated visualizations. |
| |
|
| | 3. **Spreadsheet File Evaluator** |
| | Verifies generated spreadsheet outputs with a **coarse-to-fine** strategy, combining structural checks with sampled content validation and task-specific verification. |
| |
|
| |  |
| |
|
| | *Figure 3: The design of the three types of evaluators in AIDABench.* |
| |
|
| | ## Baseline Performance |
| |
|
| | Results indicate that complex, tool-augmented document analytics remains challenging: the best-performing baseline model (**Claude-Sonnet-4.5**) achieves **59.43 pass@1** on AIDABench (see the paper for full settings, model list, and breakdowns). |
| |
|
| | ## Intended Uses |
| |
|
| | AIDABench is intended for: |
| |
|
| | - Evaluating **agents** or **tool-using LLM systems** on realistic document analytics tasks |
| | - Benchmarking end-to-end capabilities across **QA**, **file generation**, and **visualization** |
| | - Diagnosing failure modes in multi-step, multi-file reasoning over business-like data |
| |
|
| | ## Limitations |
| |
|
| | - The benchmark is designed for tool-augmented settings; purely text-only inference may underperform due to the need for code execution and file manipulation. |
| | - Automated evaluation relies on LLM judges, which introduces additional compute cost and (small) scoring variance depending on settings. |
| |
|
| | ## Citation |
| |
|
| | If you use this dataset, please cite the original paper: |
| |
|
| | ```bibtex |
| | @article{yang2026aidabench, |
| | title={AIDABENCH: AI DATA ANALYTICS BENCHMARK}, |
| | author={Yang, Yibo and Lei, Fei and Sun, Yixuan and others}, |
| | journal={arXiv preprint}, |
| | year={2026} |
| | } |
| | ``` |
| | --- |
| |
|
