AI-Driven Credit Risk Assessment for Bangladesh SMEs: Dataset and Analysis Code

1,200 SME loan records from Bangladesh, supplied by BASIC Bank Limited and Bangladesh Bank with the permission of the Ministry of Finance of Bangladesh (loans originated July 2023 to June 2025; collected July 2024 to June 2026), together with the code for (1) a centralised benchmark and (2) federated explanation experiments that test whether federated training changes the reasons given to declined applicants relative to a model trained on the pooled data, against a retraining noise floor. The research extract contains no personal identifiers or contact details. Data licence: CC BY 4.0. Code licence: MIT. Version note. The data file is identical in every version. The earlier releases (v1, 8 February 2026; v2.0, 5 June 2026), on which the IntelliSys 2026 chapter (doi:10.1007/978-3-032-32729-1_33) was prepared, described the data period as 2022 to 2023, which was wrong; version 3.0 and later state the correct dates (loans originated July 2023 to June 2025, collected July 2024 to June 2026). Version 3.0 reorganises the repository, corrects the data description, replaces mislabelled scripts and adds the explanation-drift analysis. Version 3.1.0 fits all preprocessing on training rows only, regenerates every federated result, and adds rank-graduation compliance measures and an audit on three public credit datasets. Version 3.1.1 adds federated boosted trees explained with TreeSHAP, a comparison of local and shared KernelSHAP backgrounds, differentially private FedAvg, a check against XGBoost and an audit on Lending Club loans grouped by state.