RolloutStrategy - Tinker Documentation
tinker_cookbook.rl.RolloutStrategy
class tinker_cookbook.rl.RolloutStrategy( ABC)
Controls how trajectories are collected from a group of environments.
Subclasses implement execute which receives the EnvGroupBuilder and a policy, creates envs, runs rollouts, and returns the surviving trajectories plus any error info.
Implementations must be pickleable — use @dataclass(frozen=True) with only primitive fields.
property catches_group_errors
If True, group-level errors (make_envs, compute_group_rewards) are caught and the group is skipped. If False, they propagate.
Returns: bool
execute( env_group_builder, policy)
Create envs, run rollouts, and return results.
May raise on unrecoverable errors (e.g. retry budget exhausted). The caller (do_group_rollout) handles group-level error recovery based on catches_group_errors.
Parameters:
- env_group_builder ( EnvGroupBuilder) – Builder used to create the environments for this rollout group.
- policy ( TokenCompleter) – The policy (language model) used to generate actions during rollouts.
Returns: RolloutResult – The collected trajectories, surviving environments, and any errors encountered.
Abstract method.