xdrl#

_images/xdrl-logo.png

XDRL

Hook-first interpretability for TorchRL

xdrl provides a small hook foundation for TorchRL trainer loops. Today it covers logging, evaluation, GAE, validation, and policy checkpointing; planned work explores tdhook-powered probing, steering, attribution, and representation analysis.

Get Started

Tutorials

API Reference

Key Features

_images/one.png

Trainer hooks

Attach metrics, validation, and checkpointing logic at TorchRL trainer lifecycle points.

_images/two.png

RL interpretability

Planned TensorDict-native probing, steering, attribution, and representation workflows for TorchRL trainers.