...Omniglot, Mini-ImageNet), not reinforcement learning, and includes scripts to run training and evaluation for few-shot classification. The fundamental idea is: sample a task, train on that task (inner loop), and then move the initialization parameters toward the adapted parameters (outer loop). Because Reptile is a first-order algorithm, it avoids computing second derivatives or full meta-gradients, making it computationally simpler while retaining good performance. The repo includes training scripts, dataset fetchers (Omniglot, Mini-ImageNet), and modules for defining the Reptile update logic, variables, and hyperparameters.