Trained model weights and data split for "Lightweight Deep Learning for Gastrointestinal Endoscopy Classification: A Reproducible Benchmark with Progressive Fine-Tuning and Ablation Analysis"

Trained model weights and the fixed data split accompanying the study "Lightweight Deep Learning for Gastrointestinal Endoscopy Classification: A Reproducible Benchmark with Progressive Fine-Tuning and Ablation Analysis." This archive contains:- checkpoints/ : 24 benchmark weight files (.h5) for eight lightweight backbones (EfficientNet-B0/B1, MobileNetV2, MobileNetV3-Large/Small, ShuffleNetV2, SqueezeNet 1.1, MobileViT-S), three independent runs each (seeds 42, 43, 44).- finetune/checkpoints/ : 48 fine-tuning / eight-stage ablation weight files (.h5) for EfficientNet-B1 and MobileNetV3-Small, three runs each.- data/split.json : the exact stratified train/validation/test split (seed 42; 2,804 / 600 / 602 images) used for every experiment. Models were trained on the publicly released four-class GastroEndoNet dataset (GERD, GERD-normal, polyp, polyp-normal; 4,006 endoscopic images) under one shared three-stage progressive fine-tuning schedule. Source code and analysis scripts are on GitHub (see related identifiers). Weights are provided in Keras HDF5 (.h5) format; TensorFlow 2.10, Python 3.10.