IRIS: trained model checkpoints for reliability-gated label acceptance in human-in-the-loop deep active learning

Trained model checkpoints supporting the paper IRIS: reliability-gated label acceptance and introspective instance selection for human-in-the-loop deep active learning under imperfect annotators. One checkpoint per experimental run: the final trained model after the last acquisition round, for every combination of dataset, acquisition method, oracle noise level and seed (285 runs in total). Each file stores the backbone and auxiliary head weights together with the audit trail needed to re-measure its reported accuracy independently — the indices of the examples it was trained on, the labels the (possibly noisy) synthetic oracle returned for them, the training configuration and the PyTorch version. Archives are split so a specific claim can be checked without downloading the whole set: iris-checkpoints-fmnist.tar — Fashion-MNIST, 114 models iris-checkpoints-bloodmnist.tar — BloodMNIST, 114 models iris-checkpoints-cifar10-noise{0.0,0.2}-seed{0,1,2}.tar — CIFAR-10, 57 models in six archives, one per oracle setting and seed MANIFEST.csv — every model, the archive holding it, and its recorded test accuracy SHA256SUMS.txt — checksums for all of the above Training was deterministic (cuDNN autotuning disabled, dataloader seeded), so re-running the code reproduces these numbers exactly rather than approximately. To verify: download an archive, unpack it into checkpoints/ in a clone of the code repository, and run python src/verify_checkpoint.py, which reloads each model and re-measures it on the untouched test set. Code, logs and the scripts that regenerate every number, table and figure: https://github.com/samim-reza/IRIS