Update README.md

README.md

@@ -1,6 +1,7 @@
+````md
 ---
 language:
-- en
+- zh
 license: apache-2.0
 task_categories:
 - question-answering
@@ -18,51 +19,81 @@ pretty_name: AIDABench
 
 ## Dataset Summary
 
-[cite_start]As AI-driven document understanding and processing tools become increasingly prevalent, the need for rigorous evaluation standards has grown[cite: 7]. [cite_start]**AIDABench** is a comprehensive benchmark designed for evaluating AI systems on complex Data Analytics tasks in an end-to-end manner[cite: 9]. 
+**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.
 
-[cite_start]It encompasses over 600 diverse document analytical tasks grounded in realistic scenarios, involving heterogeneous data types such as spreadsheets, databases, financial reports, and operational records[cite: 10, 11]. [cite_start]The tasks are highly challenging; even human experts require 1-2 hours per question when assisted by AI tools[cite: 12].
+![Overview of AIDABench Framework](images/figure1_overview.png)
 
-*(💡 建议在这里插入论文的 Figure 1，展示整体框架结构。)*
-![Overview of AIDABench Framework](./images/figure1_overview.png)
-[cite_start]*Figure 1: Overview of the AIDABench evaluation framework[cite: 65].*
+*Figure 1: Overview of the AIDABench evaluation framework.*
 
-## Dataset Structure
+## 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
-[cite_start]The dataset comprises three primary capability dimensions covering the end-to-end document processing pipeline[cite: 67, 199]:
-* [cite_start]**File Generation (43.3%):** Assesses data wrangling operations like filtering, format normalization, deduplication, and cross-sheet linkage[cite: 72, 199].
-* [cite_start]**Question Answering (QA) (37.5%):** Evaluates analytical operations including summation, mean computation, ranking, and trend analysis[cite: 71, 199].
-* [cite_start]**Data Visualization (19.2%):** Measures the ability to generate and adapt multiple visualization forms (bar, line, pie charts) and style customizations[cite: 73, 199].
 
-*(💡 建议在这里插入论文的 Figure 2，直观展示三种任务的输入输出示例。)*
-![Evaluation Scenarios](./images/figure2_scenarios.png)
-[cite_start]*Figure 2: Example evaluation scenarios for QA, Data Visualization, and File Generation[cite: 187, 188, 190].*
+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.
+
+![Evaluation Scenarios](images/figure2_scenarios.png)
+
+*Figure 2: Example evaluation scenarios for QA, Data Visualization, and File Generation.*
 
 ### Task Complexity
-[cite_start]Tasks are stratified by the number of expert-level reasoning steps required[cite: 201]:
-* [cite_start]**Easy (29.5%):** $\le$ 3 steps[cite: 202].
-* [cite_start]**Medium (49.4%):** 4-6 steps[cite: 202].
-* [cite_start]**Hard (21.1%):** > 7 steps[cite: 202].
-* [cite_start]**Cross-file Reasoning:** 27.4% of tasks require joint reasoning over multiple input files (up to 14 files)[cite: 205].
+
+Tasks are stratified by the number of expert-level reasoning steps required:
+
+- **Easy (29.5%)**: ≤ 3 steps  
+- **Medium (49.4%)**: 4–6 steps  
+- **Hard (21.1%)**: ≥ 7 steps  
+- **Cross-file Reasoning**: 27.4% of tasks require reasoning over multiple input files (up to 14 files).
 
 ### Data Formats
-[cite_start]Tabular files dominate the distribution (xlsx/csv account for 91.8%), complemented by DOCX and PDF formats to support mixed-type processing[cite: 200].
+
+Most inputs are tabular files (xlsx/csv dominate), complemented by **DOCX** and **PDF** formats to support mixed-type document processing.
 
 ## Evaluation Framework
 
-[cite_start]All models are evaluated under a unified, tool-augmented protocol where the model receives task instructions and files, and can execute arbitrary Python code in a sandboxed environment[cite: 211, 212, 214]. 
+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**:
 
-[cite_start]To align with the task categories, AIDABench utilizes three dedicated LLM-based evaluators[cite: 222]:
-1.  [cite_start]**QA Evaluator:** A binary judge powered by QwQ-32B[cite: 227, 326].
-2.  [cite_start]**Visualization Evaluator:** Powered by Gemini 3 Pro, scoring both correctness and readability[cite: 235, 326].
-3.  [cite_start]**Spreadsheet File Evaluator:** Powered by Claude Sonnet 4.5, utilizing a coarse-to-fine verification strategy[cite: 249, 326].
+1. **QA Evaluator**  
+   A binary judge that determines whether the produced answer matches the reference (under the benchmark’s scoring rules).
 
-*(💡 建议在这里插入论文的 Figure 3，解释自动化评测是怎么运作的。)*
-![Evaluator Design](./images/figure3_evaluators.png)
-[cite_start]*Figure 3: The design of the three types of evaluators in AIDABench[cite: 313].*
+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.
+
+![Evaluator Design](images/figure3_evaluators.png)
+
+*Figure 3: The design of the three types of evaluators in AIDABench.*
 
 ## Baseline Performance
-Results reveal that complex data analytics tasks remain a significant challenge. [cite_start]The best-performing model (Claude-Sonnet-4.5) achieved only a 59.43 pass@1 score[cite: 14, 349]. 
+
+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
 
@@ -75,6 +106,5 @@ If you use this dataset, please cite the original paper:
   journal={arXiv preprint},
   year={2026}
 }
-```
-Github Repository: https://github.com/MichaelYang-lyx/AIDABench 
----
+````
+---