SHPhoneBench: A Multi-Modal Benchmark for Second-Hand Smartphone Valuation with Heterogeneous Inspection Signals

SHPhoneBench is a multi-modal benchmark for second-hand smartphone valuation built from real recycling workflows in mainland China (January 2024 – March 2026). The full corpus contains 73,200 transaction-aligned records; a curated 15,000-sample benchmark subset carries five-tier condition grades, 18-category defect bounding boxes, cross-modal consistency labels, and differentially private transaction prices. Each record links four heterogeneous inspection signals collected during the same physical inspection:- V — multi-view exterior photographs of the device.- T — informal Chinese technician remarks (free-text).- S — structured CRM / device metadata (JSON).- IS — desktop diagnostic-tool screenshots. Tasks supported by the benchmark:- T1 — five-tier condition grading (Macro-F1, QWK).- T2 — 18-category defect detection / grounding (mAP@0.5, Grounding Acc@0.5).- T3 — binary cross-modal consistency check (F1).- T4 — calibrated price regression with 90% predictive interval (MAE in CNY, ECE). Privacy and compliance:- Faces, serial numbers, and Apple IDs are removed or masked prior to release.- List prices are released with Laplace differential privacy (epsilon = 0.1).- Collection consent is aligned with PIPL. This Zenodo record hosts the benchmark release archive, a sample manifest CSV, and a Croissant 1.1 metadata file (with MLCommons RAI 1.0 fields). Please see the accompanying paper for the full datasheet, ethics statement, and intended-use restrictions.SHPhoneBench是一款面向二手智能手机估值的多模态基准数据集，其数据采集自中国内地真实的回收业务流程（2024年1月—2026年3月）。完整语料库包含73200条交易匹配记录；经人工筛选的15000条样本基准子集涵盖五级品相等级、18类缺陷边界框标注、跨模态一致性标签，以及带差分隐私保护的交易价格。 每条记录关联同一次物理检测中采集的四类异构检测信号： - V：设备的多视角外观照片 - T：中国维修技术员的非正式自由文本备注 - S：结构化客户关系管理（Customer Relationship Management, CRM）/设备元数据（JSON格式） - IS：桌面端诊断工具截图 该基准数据集支持以下四项任务： - T1：五级品相分级（评估指标：宏F1值、二次加权卡方系数（QWK）） - T2：18类缺陷检测与定位（评估指标：平均精度均值（mean Average Precision, mAP）@0.5、定位准确率（Grounding Accuracy）@0.5） - T3：二元跨模态一致性校验（评估指标：F1值） - T4：带90%预测区间的校准价格回归（评估指标：以人民币计价的平均绝对误差（Mean Absolute Error, MAE）、期望校准误差（Expected Calibration Error, ECE）） 隐私与合规说明： 数据集发布前已对人脸信息、设备序列号以及苹果ID进行移除或掩码处理；挂牌价格采用拉普拉斯差分隐私机制（ε=0.1）进行发布；数据采集已获得符合《个人信息保护法》（PIPL）的知情同意。 本Zenodo存档包含基准数据集发布包、样本清单CSV文件以及Croissant 1.1元数据文件（内含MLCommons RAI 1.0字段）。完整数据集说明、伦理声明以及使用限制请参阅配套论文。