Class Experience

trinity/common/experience.py:105–496

@dataclass
class Experience:
    _SER_MAGIC = b"TRXP"
    _SER_VERSION = 1
    _TENSOR_FIELDS = (
        "tokens",
        "logprobs",
        "token_level_reward",
        "advantages",
        "returns",
        "action_mask",
        "chosen",
        "rejected",
        "teacher_logprobs",
        "routed_experts",
    )
    _META_FIELDS = (
        "eid",
        "reward",
        "truncate_status",
        "info",
        "metrics",
        "prompt_length",
        "response_text",
        "prompt_text",
        "messages",
        "tools",
        "chosen_messages",
        "rejected_messages",
    )

    eid: EID = field(default_factory=EID)  # Unique identifier for the experience
    tokens: Optional[Tensor] = None  # [seq_length]
    prompt_length: int = 1  # Length of the prompt in tokens, used for generating attention masks
    logprobs: Optional[Tensor] = None  # [resp_length]
    reward: Optional[float] = None
    token_level_reward: Optional[Tensor] = None  # [resp_length]
    advantages: Optional[Tensor] = None  # [resp_length]
    returns: Optional[Tensor] = None  # [resp_length]
    truncate_status: Optional[
        str
    ] = None  # The status of truncation, e.g., "prompt_truncated", "response_truncated"; Not working for openai api
    info: dict = field(
        default_factory=dict
    )  # Additional information about the experience, can also be used to store custom fields
    metrics: dict[str, float] = field(
        default_factory=dict
    )  # Metrics associated with the experience, directly used by the monitor

    # for single-turn experiences
    response_text: Optional[str] = None  # Text of the response
    prompt_text: Optional[str] = None  # Text of the prompt

    # for multi-turn experiences
    # Action mask indicates which tokens are generated by the model
    action_mask: Optional[Tensor] = None  # [resp_length]
    messages: Optional[List[dict]] = None  # List of messages
    tools: Optional[List[dict]] = None