nemo_gym.trajectory.atif_export

Strict offline conversion from Gym rollout trajectories to ATIF v1.7.

Module Contents

Classes

Name	Description
`AtifExportError`	A rollout cannot be represented by Gym’s supported ATIF profile.
`AtifExportResult`	-
`ExportAtifConfig`	Settings of `gym eval export atif`.
`_AgentGroup`	-
`_ModelCallIndex`	-

Functions

Name	Description
`_apply_turns`	-
`_attach_user_observation`	The Terminus extension: a user turn after agent output is what that step observed.
`_build_groups`	-
`_canonical_conversation_copy`	Validate a redundant turn copy and compare its normalized supported fields.
`_encoded_trajectory`	-
`_final_metrics`	-
`_index`	-
`_json_value`	-
`_live_arguments`	The live profile’s tool arguments: a string that is not a JSON object becomes `{}`, kept verbatim.
`_live_reasoning_text`	The live profile’s reading of a step’s reasoning items: lossy where a model is, never failing on it.
`_message_content`	-
`_model_call_index`	-
`_parse_arguments`	-
`_path_error`	-
`_preflight_conversation`	Reject raw fields that supported Pydantic variants would silently erase.
`_preflight_model_call_refs`	-
`_preflight_model_ref`	-
`_preflight_parts`	-
`_preflight_raw_trajectory`	-
`_preflight_training_metadata`	-
`_reasoning_text`	-
`_reject_unknown_fields`	-
`_resolve_model_call`	-
`_source_item_ids`	Return provider item identities that ATIF has no standard fields for.
`_text_content`	-
`_tool_record_extra`	-
`_transcript_items`	-
`_transcript_text`	Text content of a message or tool output; images and other parts are not representable.
`_validate_json_value`	Reject Python values that cannot round-trip through UTF-8 JSON unchanged.
`encode_atif_trajectory`	The trajectory as the UTF-8 JSON bytes Harbor reads from `trajectory.json`.
`export_rollouts_to_atif`	Validate a rollout JSONL completely, then atomically publish ATIF files and a manifest.
`gym_rollout_to_atif`	Convert one rollout row, rejecting any structure Gym cannot represent completely.
`transcript_to_atif`	Convert a live Responses API transcript into an ATIF v1.7 trajectory.

Data

TRANSCRIPT_EXPORT_PROFILE

_CONVERSATION_ADAPTER

_KNOWN_INCOMPLETE_FINISH_REASONS

_TEXT_PART_TYPES

_TRAINING_METADATA_FIELDS

API

class nemo_gym.trajectory.atif_export.AtifExportError()

Bases: ConfigError

A rollout cannot be represented by Gym’s supported ATIF profile.

class nemo_gym.trajectory.atif_export.AtifExportResult(
    output_dirpath: pathlib.Path,
    trajectory_count: int,
    manifest_fpath: pathlib.Path
)

Dataclass

manifest_fpath

Path

output_dirpath

Path

trajectory_count

int

class nemo_gym.trajectory.atif_export.ExportAtifConfig()

Bases: RunExportConfig

Settings of gym eval export atif.

agent_version

str = Field(min_length=1)

format

Literal['atif'] = 'atif'

output_dirpath

Path

session_id

str = Field(min_length=1)

nemo_gym.trajectory.atif_export.ExportAtifConfig.reject_blank_identity(
    value: str
) -> str

classmethod

class nemo_gym.trajectory.atif_export._AgentGroup(
    question: list[typing.Any],
    reasoning: list[typing.Any],
    answer: list[typing.Any],
    step: nemo_gym.trajectory.atif_v1_7.AtifStep
)

Dataclass

answer

list[Any]

question

list[Any]

reasoning

list[Any]

step

AtifStep

class nemo_gym.trajectory.atif_export._ModelCallIndex(
    calls: list[nemo_gym.trajectory.trajectory_record.TrajectoryModelCall],
    by_id: dict[str, list[int]],
    by_response: dict[tuple[str, str, str], list[int]]
)

Dataclass

by_id

dict[str, list[int]]

by_response

dict[tuple[str, str, str], list[int]]

calls

list[TrajectoryModelCall]

nemo_gym.trajectory.atif_export._apply_turns(
    trajectory: nemo_gym.trajectory.trajectory_record.TrajectoryRecord,
    groups: list[nemo_gym.trajectory.atif_export._AgentGroup]
) -> set[int]

nemo_gym.trajectory.atif_export._attach_user_observation(
    step: nemo_gym.trajectory.atif_v1_7.AtifStep,
    item: typing.Any,
    path: str
) -> None

The Terminus extension: a user turn after agent output is what that step observed.

nemo_gym.trajectory.atif_export._build_groups(
    trajectory: nemo_gym.trajectory.trajectory_record.TrajectoryRecord,
    invocation: typing.Any
) -> tuple[list[nemo_gym.trajectory.atif_v1_7.AtifStep], list[nemo_gym.trajectory.atif_export._AgentGroup]]

nemo_gym.trajectory.atif_export._canonical_conversation_copy(
    value: typing.Any,
    path: str
) -> typing.Any

Validate a redundant turn copy and compare its normalized supported fields.

nemo_gym.trajectory.atif_export._encoded_trajectory(
    trajectory: nemo_gym.trajectory.atif_v1_7.AtifTrajectoryV1_7
) -> bytes

nemo_gym.trajectory.atif_export._final_metrics(
    groups: list[nemo_gym.trajectory.atif_export._AgentGroup],
    total_steps: int
) -> nemo_gym.trajectory.atif_v1_7.AtifFinalMetrics

nemo_gym.trajectory.atif_export._index(
    row: dict[str, typing.Any],
    key: str,
    path: str
) -> int

nemo_gym.trajectory.atif_export._json_value(
    value: typing.Any
) -> typing.Any

nemo_gym.trajectory.atif_export._live_arguments(
    call: nemo_gym.core.responses_api_types.NeMoGymResponseFunctionToolCall,
    reasons: set[str]
) -> tuple[dict[str, typing.Any], dict[str, typing.Any] | None]

The live profile’s tool arguments: a string that is not a JSON object becomes {}, kept verbatim.

A harness that survives a malformed call (it hands the model a tool error and goes on) produced a gradable episode; the original string travels in the call’s extra and the trajectory is partial.

nemo_gym.trajectory.atif_export._live_reasoning_text(
    items: list[nemo_gym.core.responses_api_types.NeMoGymResponseReasoningItem],
    path: str,
    reasons: set[str]
) -> str | None

The live profile’s reading of a step’s reasoning items: lossy where a model is, never failing on it.

Empty and encrypted reasoning (OpenAI reasoning models answer with summary: [] unless a summary is requested) export without reasoning_content; several segments, from several items or one item’s summary list (one <think> block each), join with newlines. Each departure from the strict mapping is named in reasons, so the trajectory says it is partial. Only a segment without text (a converter defect) still raises.

nemo_gym.trajectory.atif_export._message_content(
    item: typing.Any,
    path: str
) -> nemo_gym.trajectory.atif_v1_7.AtifContent

nemo_gym.trajectory.atif_export._model_call_index(
    calls: list[nemo_gym.trajectory.trajectory_record.TrajectoryModelCall]
) -> nemo_gym.trajectory.atif_export._ModelCallIndex

nemo_gym.trajectory.atif_export._parse_arguments(
    call: nemo_gym.core.responses_api_types.NeMoGymResponseFunctionToolCall,
    path: str
) -> dict[str, typing.Any]

nemo_gym.trajectory.atif_export._path_error(
    path: str,
    detail: str
) -> nemo_gym.trajectory.atif_export.AtifExportError

nemo_gym.trajectory.atif_export._preflight_conversation(
    conversation: typing.Any,
    path: str
) -> None

Reject raw fields that supported Pydantic variants would silently erase.

nemo_gym.trajectory.atif_export._preflight_model_call_refs(
    value: typing.Any,
    path: str
) -> None

nemo_gym.trajectory.atif_export._preflight_model_ref(
    value: typing.Any,
    path: str
) -> None

nemo_gym.trajectory.atif_export._preflight_parts(
    value: typing.Any,
    path: str,
    supported_fields: dict[str, frozenset[str]]
) -> None

nemo_gym.trajectory.atif_export._preflight_raw_trajectory(
    value: typing.Any
) -> None

nemo_gym.trajectory.atif_export._preflight_training_metadata(
    item: dict[typing.Any, typing.Any],
    path: str
) -> None

nemo_gym.trajectory.atif_export._reasoning_text(
    items: list[nemo_gym.core.responses_api_types.NeMoGymResponseReasoningItem],
    path: str
) -> str | None

nemo_gym.trajectory.atif_export._reject_unknown_fields(
    value: dict[typing.Any, typing.Any],
    allowed: set[str] | frozenset[str],
    path: str
) -> None

nemo_gym.trajectory.atif_export._resolve_model_call(
    ref: nemo_gym.trajectory.trajectory_record.ModelCallRef,
    calls: nemo_gym.trajectory.atif_export._ModelCallIndex,
    path: str
) -> tuple[int, nemo_gym.trajectory.trajectory_record.TrajectoryModelCall]

nemo_gym.trajectory.atif_export._source_item_ids(
    group: nemo_gym.trajectory.atif_export._AgentGroup
) -> dict[str, typing.Any]

Return provider item identities that ATIF has no standard fields for.

nemo_gym.trajectory.atif_export._text_content(
    value: typing.Any,
    path: str
) -> nemo_gym.trajectory.atif_v1_7.AtifContent

nemo_gym.trajectory.atif_export._tool_record_extra(
    record: typing.Any,
    source_item_id: str | None = None
) -> dict[str, typing.Any] | None

nemo_gym.trajectory.atif_export._transcript_items(
    items: typing.Any,
    path: str
) -> list[typing.Any]

nemo_gym.trajectory.atif_export._transcript_text(
    value: typing.Any,
    path: str,
    reasons: set[str] | None = None
) -> nemo_gym.trajectory.atif_v1_7.AtifContent

Text content of a message or tool output; images and other parts are not representable.

With reasons (the model’s own output), a refusal part exports as its text and is noted there: the model refused, the harness went on, and the episode is still graded, as Harbor grades a refused Terminus turn (Harbor keeps an empty step; the text is kept here). Without it, a refusal is rejected like any other non-text part.

nemo_gym.trajectory.atif_export._validate_json_value(
    value: typing.Any,
    path: str
) -> None

Reject Python values that cannot round-trip through UTF-8 JSON unchanged.

nemo_gym.trajectory.atif_export.encode_atif_trajectory(
    trajectory: nemo_gym.trajectory.atif_v1_7.AtifTrajectoryV1_7
) -> bytes

The trajectory as the UTF-8 JSON bytes Harbor reads from trajectory.json.

nemo_gym.trajectory.atif_export.export_rollouts_to_atif(
    config: nemo_gym.trajectory.atif_export.ExportAtifConfig
) -> nemo_gym.trajectory.atif_export.AtifExportResult

Validate a rollout JSONL completely, then atomically publish ATIF files and a manifest.

nemo_gym.trajectory.atif_export.gym_rollout_to_atif(
    rollout: dict[str, typing.Any],
    session_id: str,
    agent_version: str
) -> nemo_gym.trajectory.atif_v1_7.AtifTrajectoryV1_7

Convert one rollout row, rejecting any structure Gym cannot represent completely.

nemo_gym.trajectory.atif_export.transcript_to_atif(
    input_items: typing.Any,
    output_items: typing.Any,
    session_id: str,
    agent_name: str,
    agent_version: str,
    model_name: str | None = None,
    usage: collections.abc.Mapping[str, typing.Any] | None = None
) -> nemo_gym.trajectory.atif_v1_7.AtifTrajectoryV1_7

Convert a live Responses API transcript into an ATIF v1.7 trajectory.

input_items is the task’s responses_create_params.input (a string or item list) and output_items the agent’s response.output. Shapes a model causes (reasoning without an answer when it ran out of tokens, several reasoning segments, empty or encrypted reasoning, arguments that are not a JSON object) export with the departure named under extra.nemo_gym.conversion (status: partial, reasons); raises AtifExportError only for what the mapping cannot represent at all (a copied assistant turn, a later system turn, a result for no call or a second function_call_output for the same call_id, item types ATIF has no place for).

nemo_gym.trajectory.atif_export.TRANSCRIPT_EXPORT_PROFILE = 'ng-transcript-to-atif-v1'

nemo_gym.trajectory.atif_export._CONVERSATION_ADAPTER = TypeAdapter(list[NeMoGymResponseInputItem])

nemo_gym.trajectory.atif_export._KNOWN_INCOMPLETE_FINISH_REASONS = frozenset({'aborted', 'canceled', 'cancelled', 'content_filter', 'failed', 'inco...

nemo_gym.trajectory.atif_export._TEXT_PART_TYPES = frozenset({'input_text', 'output_text', 'text'})

nemo_gym.trajectory.atif_export._TRAINING_METADATA_FIELDS = frozenset({'prompt_token_ids', 'generation_token_ids', 'generation_log_probs', '...