ESGBench
收藏资源简介:
ESGBench是由纽约梅隆银行与联邦快递联合构建的可解释ESG问答基准数据集,聚焦企业可持续发展报告分析。该数据集包含119个高质量问答对,涵盖环境、社会、治理三大维度,数据源自12份跨国企业的TCFD报告PDF文档,通过自动化流程生成带原文证据的问答数据。采用检索增强生成技术构建基准测试框架,主要应用于数字金融领域的ESG指标提取、可持续投资决策支持等场景,旨在解决非结构化ESG报告难以量化分析的行业痛点。
ESGBench is an interpretable ESG question-answering benchmark dataset jointly developed by Bank of New York Mellon and FedEx, focusing on enterprise sustainability report analysis. This dataset contains 119 high-quality question-answer pairs covering the three core dimensions of Environment, Social and Governance (ESG). It is sourced from 12 PDF documents of TCFD reports issued by multinational corporations, and all question-answer pairs with supporting original text evidence are generated via automated workflows. The benchmark test framework is constructed using Retrieval-Augmented Generation (RAG) technology, which is primarily applied in scenarios including ESG indicator extraction and sustainable investment decision support within the digital finance domain. This work aims to address the industry pain point that unstructured ESG reports pose challenges for quantitative analysis.
ESGBench 数据集概述
数据集简介
ESGBench 是一个可解释的 ESG 问答基准数据集,通过自动化流程从 ESG/TCFD PDF 文档生成具有证据支持的问答对。
核心功能
- 收集 ESG/TCFD PDF 文档
- 构建可搜索索引和表格缓存
- 自动生成基于证据的 ESG 问答对
- 运行 RAG 基线(可选)
- 评估预测结果(EM/F1/数值准确率/Recall@K)
数据生成流程
1. 种子文档准备
编辑 data/docs_seed.csv 文件(UTF-8编码,需要表头):
company,year,url,doc_type,country,industry,source Apple Inc,2024,https://…/apple-2024-esg.pdf,ESG,US,Technology,manual
2. 文档处理
运行 python -m scripts.ingest_catalog data/docs_seed.csv:
- 下载 PDF 到 pdfs/ 目录
- 记录到 data/esgbench_document_information.jsonl
3. 索引构建
运行 python -m scripts.build_index:
- 生成 cache/chunks.json(文本块)
- 生成 cache/<DOC>_tables.json(解析的表格数据)
4. 问答对生成
运行 python -m scripts.generate_qas_from_chunks:
- 输出到 data/esgbench_open_source.jsonl
- 支持去重和防护机制
数据格式
问答对示例
json { "company": "Apple Inc", "doc_name": "APPLE_INC_2024_ESG", "category": "Environmental", "kpi_name": "Scope 2 (market-based)", "question": "What are Apple Incs Scope 2 (market-based) emissions in 2024?", "answer": "1,234,567 tCO2e", "evidence": [{ "evidence_text": "… Scope 2 (market-based) were 1,234,567 tCO2e in 2024 …", "evidence_page_num": 39, "evidence_doc_name": "APPLE_INC_2024_ESG" }] }
评估指标
- 精确匹配(Exact Match)
- 字符串 F1 分数
- 数值准确率(±2% 容差,支持单位识别)
- 检索召回率@K
- 分类别准确率
技术环境
- Python 3.10–3.12
- 需要 OpenAI API 密钥
- 支持 macOS(需要安装 Ghostscript 和 Java)
目录结构
esgbench/ data/ # 种子数据和生成的标注/预测 pdfs/ # 下载的 PDF 文档 cache/ # 块、表格、嵌入向量 scripts/ # CLI 入口点 src/esgbench/ # 可导入的库代码
环境变量
- OPENAI_API_KEY(必需)
- LLM_MODEL(默认 gpt-5-mini)
- EMB_MODEL(默认 text-embedding-3-large)
- RETRIEVE_K(默认 5)
- PASSAGE_CHARS(默认 900)




