ML Benchmark Data for An open-source machine-learning benchmark for Raman species identification of wound-associated bacteria and polymicrobial mixture quantification

With this dataset, we benchmarked random forest, XGBoost, convolutional neural network (CNN), and Transformer classifiers using Raman spectra from six wound-associated bacterial species and five Staphylococcus aureus/Pseudomonas aeruginosa mixture ratios. Unsupervised dimensionality reduction showed substantial overlap among classes. The CNN achieved the highest clean-test accuracy for species identification (98.74%) and mixture-ratio classification (97.04%) and remained stable after synthetic white-noise perturbation; all models showed less than 0.5 percentage-point degradation, and noise-augmented retraining improved performance. These results establish a reproducible benchmark for supervised analysis of Raman bacterial spectra and show that mixture ratios can be discriminated in laboratory-prepared samples. The dataset, trained models, and preprocessing code are openly available to support validation on more complex mixtures and clinical wound specimens and can be found here: https://github.com/Tissue-Engineering-BioImaging-Lab/ML-Benchmark-for-Species-Identification-and-Polymicrobial-Mixture-Quantification/tree/main