Hacker News 社区展示了开源工具集 Claimseed v0.6.0,该工具旨在生成带有植入错误和答案键的合成美国 Medicare 医疗账单文档(如摘要通知、明细账单等),为文档提取器、账单审核工具和 AI 代理提供安全的测试环境,以避免接触真实患者隐私 1。最新版本发布了包含 1,022 份盲测文档的标准数据集,具体包括 260 份 Medicare Summary Notices、250 份 itemized bills、256 份 PPO EOBs 以及 256 份 HMO EOBs 1。所有生成的姓名、标识符、地址和金额均由种子生成,确保没有真实文档数据被使用或衍生 1。每页文档均通过标本带、对角线水印和 PDF 元数据标记为合成样本,且无法通过 CLI 或代码关闭这些标记 1。
在技术实现与评估机制方面,Claimseed 引入了“盲密钥”机制 (--blind-keys),将答案键与文档分离存储,防止测试者作弊或模型训练数据污染 1。同一种子在同一 Chrome 构建(Linux CI)下渲染的 PDF 字节完全相同,以确保可复现性 1。系统支持 12 种布局变体 (b01-b12, p01-p12, h01-h12) 以测试工具对版式变化的鲁棒性,并支持降级图像模拟以及与 X12 835 汇款单的配套测试 1。此外,开发者推出了 MCP (Model Context Protocol) 服务器 (@claimseed/mcp),允许 AI 代理直接调用渲染和验证功能 1。配套的评分系统计算字段准确率、错误召回率、精确率和误报率,其中清洁场景专门用于检测误报 1。
Claimseed, an open-source toolset designed to generate synthetic US Medicare billing documents such as summary notices and itemized bills, has released version 0.6.0 on Hacker News. 1 The update introduces a standard dataset containing 1,022 blind-test documents, specifically comprising 260 Medicare Summary Notices, 250 itemized bills, 256 PPO Explanation of Benefits (EOBs), and 256 HMO EOBs. 1 The tool aims to provide a safe testing environment for document extractors, billing audit tools, and AI agents by avoiding contact with real patient privacy data. 1
All names, identifiers, addresses, and monetary amounts in the generated documents are created via seeds to ensure no real-world document data is used or derived. 1 To prevent test cheating and contamination of model training data, the release includes a "blind keys" mechanism (--blind-keys) that separates answer keys from the documents themselves. 1 Each page is permanently marked as a synthetic sample through specimen bands, diagonal watermarks, and PDF metadata; these markers cannot be disabled via command-line interface or code. 1
The new version supports 12 layout variants (b01-b12, p01-p12, h01-h12) to test the robustness of tools against format variations. 1 It also features degraded image simulation and paired testing with X12 835 remittance advices. 1 For reproducibility, PDFs rendered from the same seed on identical Chrome builds (Linux CI) produce byte-identical results. 1 The tool provides an MCP (Model Context Protocol) server (@claimseed/mcp), allowing AI agents to directly call rendering and validation functions. 1 Its scoring system calculates field accuracy, error recall, precision, and false positive rates, with clean scenarios specifically included to detect false positives. 1
评论
还没有评论,欢迎留下第一条。