HF RL Explorer

Add support for saving and loading custom Oumi models

Add support for saving and loading custom Oumi models: a task in LegoFlow-SWE (Harbor dataset). Currently, custom registry-based models (e.g., MLPEncoder, CNNClassifier, NanoGPT) cannot be saved or reloaded after training. This makes it impossible to reuse a pretrained custom model for inference…

The task

Currently, custom registry-based models (e.g., MLPEncoder, CNNClassifier, NanoGPT) cannot be saved or reloaded after training. This makes it impossible to reuse a pretrained custom model for inference, finetuning, or evaluation without retraining from scratch.

Part of Lego-X/LegoFlow-SWE.