Turn
One assistant turn of a rollout.
prompt_tokens is the full rendered context the policy saw for this turn
(system + prior turns + the latest observation); completion_tokens /
logprobs are the sampled response. A single-turn rollout has exactly one.
logprobs is optional - on-policy tinker rollouts populate it (training
needs it), but eval / closed-API rollouts often have none. Defaults to empty.
Attributes
attributeprompt_tokenslist[int]attributecompletion_tokenslist[int]attributelogprobslist[float]= field(default_factory=list)attributetextstr= ''Functions
func__init__(self, prompt_tokens, completion_tokens, logprobs=list(), text='') -> Noneparamselfparamprompt_tokenslist[int]paramcompletion_tokenslist[int]paramlogprobslist[float]= list()paramtextstr= ''Returns
None