trinity.common.experience module

Experience Class.

class trinity.common.experience.EID(batch: int | str = '', task: int | str = '', run: int = 0, step: int = 0, suffix: str = <factory>)

基类：object

Experience ID class to uniquely identify an experience.

To enable the full functionality of the experience grouping, user should manually set the run and step fields in custom workflows.

batch: int | str = ''

task: int | str = ''

run: int = 0

step: int = 0

suffix: str

property uid: str

An unique identifier for the experience.

property sid: str

Step ID of the experience.

For example, experiences generated by all runs of a same task at the same step will have the same sid.

property rid: str

Run ID of the experience.

For example, experiences generated by one run of a task at all steps will have the same run_id.

property tid: str

Task ID for the experience.

For example, experiences generated by a all run of a same task in GRPO-like algorithms will have the same tid.

to_dict() → dict

Convert the EID to a dictionary.

__init__(batch: int | str = '', task: int | str = '', run: int = 0, step: int = 0, suffix: str = <factory>) → None

class trinity.common.experience.CustomField(source_field: str, destination_field: str, data_type: dtype)

基类：object

Custom field for Experiences.

This is used to store additional information into the Experiences class.

source_field: str

destination_field: str

data_type: dtype

__init__(source_field: str, destination_field: str, data_type: dtype) → None

class trinity.common.experience.Experience(*, eid=None, tokens, logprobs=None, reward=None, token_level_reward=None, advantages=None, returns=None, truncate_status=None, info=None, metrics=None, prompt_length=1, response_text=None, prompt_text=None, action_mask=None, messages=None, tools=None, chosen=None, rejected=None, chosen_messages=None, rejected_messages=None, multi_modal_inputs=None, teacher_logprobs=None, custom_fields=None)

基类：object

__init__(*, eid=None, tokens, logprobs=None, reward=None, token_level_reward=None, advantages=None, returns=None, truncate_status=None, info=None, metrics=None, prompt_length=1, response_text=None, prompt_text=None, action_mask=None, messages=None, tools=None, chosen=None, rejected=None, chosen_messages=None, rejected_messages=None, multi_modal_inputs=None, teacher_logprobs=None, custom_fields=None)

eid: EID

reward: float | None = None

token_level_reward: Tensor | None = None

advantages: Tensor | None = None

returns: Tensor | None = None

info: dict

metrics: dict[str, float]

truncate_status: str | None = None

prompt_length: int = 1

response_text: str | None = None

prompt_text: str | None = None

messages: List[dict] | None = None

tools: List[dict] | None = None

chosen_messages: List[dict] | None = None

rejected_messages: List[dict] | None = None

multi_modal_inputs: Dict[str, Tensor] | None = None

tokens: Tensor | None = None

logprobs: Tensor | None = None

action_mask: Tensor | None = None

chosen: Tensor | None = None

rejected: Tensor | None = None

teacher_logprobs: Tensor | None = None

custom_fields: List[CustomField]

serialize() → bytes

Serialize the experience to bytes.

classmethod deserialize(data: bytes) → Experience

classmethod serialize_many(experiences: List[Experience]) → bytes

Serialize a list of experiences into a compact bytes payload.

Tensor fields are packed with safetensors while non-tensor fields are packed as metadata via pickle.

classmethod deserialize_many(data: bytes) → List[Experience]

Deserialize bytes into a list of experiences.

Supports both new batched payloads and legacy single-experience pickle payloads.

to_dict() → dict

Convert the experience to a dictionary.

trinity.common.experience.split_dpo_experience_to_single_turn(experiences: List[Experience]) → List[Experience]

trinity.common.experience.gather_token_ids(experiences, max_prompt_length: int, max_response_length: int, pad_token_id: int) → Tensor

trinity.common.experience.gather_action_masks(experiences, max_response_length: int) → Tensor

trinity.common.experience.gather_attention_masks(experiences, max_prompt_length: int, max_response_length: int) → Tensor

trinity.common.experience.gather_response_attrs(experiences, attr_name: str, max_response_length: int, pad_value: int = 0) → Tensor

trinity.common.experience.gather_multi_modal_inputs(experiences) → Dict[str, Tensor]

trinity.common.experience.group_by(experiences: List[Experience], id_type: Literal['task', 'run', 'step']) → Dict[str, List[Experience]]

Group experiences by ID.

trinity.common.experience.to_hf_datasets(experiences: list[Experience]) → Dataset

Convert a list of Experience objects to a HuggingFace Dataset, preserving all fields.

trinity.common.experience.from_hf_datasets(dataset: Dataset) → List[Experience]

Convert a HuggingFace Dataset back to a list of Experience objects.