Copy as Markdown View as Markdown

tinker_cookbook.rl.assemble_training_data

Generated from tinker-cookbook at commit 8486b8d. Source links point at that snapshot.

tinker_cookbook.rl.assemble_training_data(trajectory_groups_P, advantages_P, include_provenance)

Convert trajectories to training data format.

Parameters:

• trajectory_groups_P (list[TrajectoryGroup]) – Groups of trajectories to convert into training datums.
• advantages_P (list[torch.Tensor]) – Per-group advantage tensors, one per trajectory group, as returned by compute_advantages.
• include_provenance (bool) – Passed to trajectory_to_data. Defaults to True.

Returns: tuple[list[tinker.Datum], list[dict[str, int]]] – A flat list of training datums and a parallel list of metadata dicts mapping each datum back to its group_idx and traj_idx.