reference/generated/torchrl.trainers.algorithms.configs.transforms.KLRewardTransformConfig

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

torchrl.trainers.algorithms.configs.transforms.KLRewardTransformConfig

class torchrl.trainers.algorithms.configs.transforms.KLRewardTransformConfig(ref_model: Any = None, ref_model_factory: Any = None, coef: Any = 1.0, in_keys: list[str] | None = None, out_keys: list[str] | None = None, log_prob_key: Any = ('log_probs', 'full'), device: Any = None, add_to_reward: bool = True, tokenizer: Any = None, assistant_only: bool = True, padding_side: str = 'left', use_ray_service: bool = False, _target_: str = 'torchrl.envs.llm.transforms.kl.KLRewardTransform')

Hydra configuration for KLRewardTransform.

Every kwarg accepted by KLRewardTransform.__init__ is exposed as a field here. Instantiating without ref_model or ref_model_factory still fails at runtime because the class requires one of them.