Rethinking Long-tailed Dataset Distillation: A Uni-Level Framework with Unbiased Recovery and Relabeling

The paper has been accepted as AAAI 2026 oral.

Expert Model Training

For example, to train on CIFAR-10-LT, run:

cd expert
python main.py --dataset cifar10 -a convnet --num_classes 10 --imbanlance_rate 0.01--epochs 200 -b 64 --q 0.8 --gamma1 1

Distilled Image Initialization

To run a specific experiment, use the corresponding script. For example:

cd initial
sh scripts/cifar10_10ipc_conv3_to_conv3_cr5.sh

Unbiased Recovery

CIFAR-10-LT

Adjust the paths and hyperparameters in recover_cifar10/recover.sh and recover_cifar10/recover.py, then run:

cd recover_cifar10
sh recover.sh

CIFAR-100-LT

Adjust the paths and hyperparameters in recover_cifar100/recover.sh and recover_cifar100/recover.py, then run:

cd recover_cifar100
sh recover.sh

Tiny-ImageNet-LT

Change imbalance factor in recover_tiny/tiny_in_dataset.py, adjust the paths and hyperparameters in recover_tiny/recover.sh and recover_tiny/recover.py , then run:

cd recover_tiny
sh recover.sh

ImageNet-LT

Adjust the paths and hyperparameters in recover_1k/recover.sh and recover_1k/recover.py, then run:

cd recover_1k
sh recover.sh

Unbiased Relabeling and Student Training

CIFAR-10-LT

Adjust the paths and hyperparameters in train_cifar10/train.sh and train_cifar10/direct_train.py, then run:

cd train_cifar10
sh train.sh

CIFAR-100-LT

Adjust the paths and hyperparameters in train_cifar100/train.sh and train_cifar100/direct_train.py, then run:

cd train_cifar100
sh train.sh

Tiny-ImageNet-LT

Change imbalance factor in train_tiny/tiny_in_dataset.py, adjust the paths and hyperparameters in train_tiny/train.sh and train_tiny/direct_train.py, then run:

cd train_tiny
sh train.sh

ImageNet-LT

Adjust the paths and hyperparameters in train_1k/train.sh and train_1k/direct_train.py, then run:

cd train_1k
sh train.sh