DianJin/DianJin-Fin-PRM-Data

--- language: - zh license: apache-2.0 size_categories: - 1K<n<10K task_categories: - question-answering - text-generation tags: - finance - process-reward-model - PRM - Chinese - evaluation dataset_info: features: - name: 名称 dtype: string - name: 科目 dtype: string - name: 章节 dtype: string - name: task dtype: string - name: question dtype: string - name: choices dtype: string - name: answer dtype: string - name: analysis dtype: string - name: analysis_length dtype: int64 - name: trace dtype: string - name: final_answer dtype: string - name: knowlegde_coverage_response dtype: string - name: knowledge_coverage_score dtype: float64 - name: coverage_knowledge_points dtype: string - name: importance_labels dtype: string - name: analysis_knowledge dtype: string - name: knowledge_accuracy dtype: string - name: quality_score dtype: string - name: model_answer dtype: string - name: accuracy_score dtype: string - name: step_scores dtype: string - name: step_labels dtype: string - name: trajectory_label dtype: int64 splits: - name: train num_examples: 4969 --- # DianJin-Fin-PRM Dataset ## Overview DianJin-Fin-PRM is a Chinese financial domain **Process Reward Model (PRM)** training dataset. It contains 4,969 samples of financial exam questions with step-by-step reasoning traces and multi-dimensional quality annotations. ## Dataset Structure | Field | Type | Description | |---|---|---| | `名称` | string | Exam name (e.g., 初级经济师) | | `科目` | string | Subject (e.g., 金融实务) | | `章节` | string | Chapter | | `task` | string | Question type (e.g., 单项选择题) | | `question` | string | Question text | | `choices` | string | Answer choices (JSON dict) | | `answer` | string | Ground truth answer | | `analysis` | string | Reference analysis | | `analysis_length` | int | Length of reference analysis | | `trace` | string | Step-by-step reasoning trace | | `final_answer` | string | Model's final answer with reasoning | | `knowlegde_coverage_response` | string | Knowledge coverage evaluation response | | `knowledge_coverage_score` | float | Knowledge coverage score (0-1) | | `coverage_knowledge_points` | string | Covered knowledge points (JSON list) | | `importance_labels` | string | Step importance labels (JSON list) | | `analysis_knowledge` | string | Knowledge points from analysis (JSON list) | | `knowledge_accuracy` | string | Per-step knowledge accuracy (JSON list) | | `quality_score` | string | Per-step quality scores (JSON list of dicts with logical_soundness, step_correctness, target_progression) | | `model_answer` | string | Model's extracted answer | | `accuracy_score` | string | Per-step accuracy scores (JSON list) | | `step_scores` | string | Aggregated per-step scores (JSON list) | | `step_labels` | string | Per-step binary labels (JSON list) | | `trajectory_label` | int | Overall trajectory label (1=correct, 0=incorrect) | ## Usage ```python from datasets import load_dataset dataset = load_dataset("DianJin/DianJin-Fin-PRM-Data") ``` ## Citation If you use this dataset, please cite: ```bibtex @misc{dianjin-fin-prm, title={DianJin-Fin-PRM: A Chinese Financial Process Reward Model Dataset}, author={DianJin Team}, year={2025} } ``` ## License Apache License 2.0--- 语言: - zh 许可证: apache-2.0 规模类别: - 1K<n<10K 任务类别: - 问答 - 文本生成 标签: - 金融 - 过程奖励模型（Process Reward Model，PRM） - 中文 - 评估 数据集信息: 特征: - 名称: dtype: 字符串 - 科目: dtype: 字符串 - 章节: dtype: 字符串 - task: dtype: 字符串 - question: dtype: 字符串 - choices: dtype: 字符串 - answer: dtype: 字符串 - analysis: dtype: 字符串 - analysis_length: dtype: 整数 - trace: dtype: 字符串 - final_answer: dtype: 字符串 - knowlegde_coverage_response: dtype: 字符串 - knowledge_coverage_score: dtype: 浮点数 - coverage_knowledge_points: dtype: 字符串 - importance_labels: dtype: 字符串 - analysis_knowledge: dtype: 字符串 - knowledge_accuracy: dtype: 字符串 - quality_score: dtype: 字符串 - model_answer: dtype: 字符串 - accuracy_score: dtype: 字符串 - step_scores: dtype: 字符串 - step_labels: dtype: 字符串 - trajectory_label: dtype: 整数 划分集: - 名称: train 样本数: 4969 --- # DianJin-Fin-PRM 数据集 ## 概览 DianJin-Fin-PRM 是一款面向中文金融领域的**过程奖励模型（Process Reward Model，PRM）**训练数据集。该数据集包含4969道金融考题样本，附带分步推理轨迹与多维度质量标注。 ## 数据集结构 | 字段 | 类型 | 描述 | |---|---|---| | `名称` | 字符串 | 考试名称（例如：初级经济师） | | `科目` | 字符串 | 科目（例如：金融实务） | | `章节` | 字符串 | 所属章节 | | `task` | 字符串 | 题型（例如：单项选择题） | | `question` | 字符串 | 试题文本 | | `choices` | 字符串 | 答案选项（JSON格式字典） | | `answer` | 字符串 | 标准答案 | | `analysis` | 字符串 | 参考解析 | | `analysis_length` | 整数 | 参考解析长度 | | `trace` | 字符串 | 分步推理轨迹 | | `final_answer` | 字符串 | 带完整推理过程的模型最终答案 | | `knowlegde_coverage_response` | 字符串 | 知识覆盖度评估响应 | | `knowledge_coverage_score` | 浮点数 | 知识覆盖度得分（取值范围0-1） | | `coverage_knowledge_points` | 字符串 | 覆盖知识点（JSON格式列表） | | `importance_labels` | 字符串 | 步骤重要性标签（JSON格式列表） | | `analysis_knowledge` | 字符串 | 解析中涵盖的知识点（JSON格式列表） | | `knowledge_accuracy` | 字符串 | 单步知识准确率（JSON格式列表） | | `quality_score` | 字符串 | 单步质量得分（JSON格式字典列表，包含逻辑合理性、步骤正确性、目标推进度三个维度） | | `model_answer` | 字符串 | 模型提取的最终答案 | | `accuracy_score` | 字符串 | 单步准确率得分（JSON格式列表） | | `step_scores` | 字符串 | 聚合单步得分（JSON格式列表） | | `step_labels` | 字符串 | 单步二元分类标签（JSON格式列表） | | `trajectory_label` | 整数 | 整体推理轨迹标签（1代表正确，0代表错误） | ## 使用方法 python from datasets import load_dataset dataset = load_dataset("DianJin/DianJin-Fin-PRM-Data") ## 引用声明 若您使用本数据集，请引用以下文献： bibtex @misc{dianjin-fin-prm, title={DianJin-Fin-PRM: A Chinese Financial Process Reward Model Dataset}, author={DianJin Team}, year={2025} } ## 许可证 Apache 2.0 许可证