tinker_cookbook.rl.compute_advantages

tinker_cookbook.rl.compute_advantages( trajectory_groups_P)

Compute advantages for each trajectory, centered within groups.

Parameters:

• trajectory_groups_P ( _list[ TrajectoryGroup – Groups of trajectories, where each group's rewards are centered independently.

Returns: list[torch.Tensor] – Per-group advantage tensors of shape (G,), where G is the number of trajectories in each group.

Referenced by

• tinker_cookbook.rl.assemble_training_data