| configs: | |
| - config_name: QA | |
| data_files: QA.json | |
| features: | |
| - name: question | |
| dtype: string | |
| - name: ground_truth | |
| dtype: string | |
| - config_name: Documents | |
| data_files: Documents.json | |
| features: | |
| - name: id | |
| dtype: string | |
| - name: content | |
| dtype: string | |
| # Dataset Structure | |
| This dataset contains two subsets: | |
| - **QA**: Question-answer pairs (Single-turn) | |
| - `question` (string): The question text. | |
| - `ground_truth` (string): The reference response. | |
| - **Documents**: Document contents | |
| - `id` (string): Unique document identifier. | |
| - `content` (string): The document text content. | |
| ## Data Construction | |
| The data is constructed using LLM-Generated Data; | |
| Questions are synthetically generated by a Large Language Model based on provided reference documents. The model is instructed to create queries whose answers exist within the given context. The reference document text is stored in the **Documents** subset. | |
| ## Source | |
| Paper abstracts. | |
| ## Review Process | |
| All data undergoes a manual human review process. Problematic samples are directly removed or modified while preserving their original intent. Reviewers may also use automated tools to assist in this process. | |
| | # | Criterion | Description | | |
| |:-:|:--|:--| | |
| | 1 | Human-Centric Completeness and Meaningfulness | The generated question must be complete, unambiguous, and fully meaningful from a human evaluator's perspective. Vague or robotic-sounding questions are removed. | | |
| *Example (Rejected): "Based on the above clause, what is the ruling?" (lacks clear reference to the specific clause).* | |
| *Example (Accepted): "Based on Clause A of Article 3, what is the ruling regarding unilateral termination?"* |