nat.plugins.eval.runtime.atif_adapter

ATIF adapter utilities for eval runtime ingress.

This module provides a single-conversion adapter layer from EvalInputItem trajectory data to ATIFTrajectory objects. Runtime code uses this to avoid per-evaluator conversion and to keep ATIF as the canonical internal trace shape.

Classes

EvalAtifAdapter

Build and cache ATIF trajectories for eval items.

Module Contents

class EvalAtifAdapter(

converter: nat.utils.atif_converter.IntermediateStepToATIFConverter | None = None,

)

Build and cache ATIF trajectories for eval items.

_converter

_cache: dict[str, nat.atif.ATIFTrajectory]

static _cache_key(item_id: Any) → str

_coerce_trajectory(value: Any) → nat.atif.ATIFTrajectory

get_trajectory(

item: nat.data_models.evaluator.EvalInputItem,

prebuilt: nat.atif.ATIFTrajectory | collections.abc.Mapping[str, Any] | None = None,

) → nat.atif.ATIFTrajectory

Return cached ATIF trajectory for an eval item, converting at most once.

_ensure_cache(

eval_input: nat.data_models.evaluator.EvalInput,

prebuilt_trajectories: collections.abc.Mapping[str, nat.atif.ATIFTrajectory | collections.abc.Mapping[str, Any]] | None = None,

) → None

Populate cache for all eval items.

build_samples(

eval_input: nat.data_models.evaluator.EvalInput,

prebuilt_trajectories: collections.abc.Mapping[str, nat.atif.ATIFTrajectory | collections.abc.Mapping[str, Any]] | None = None,

) → nat.plugins.eval.evaluator.atif_evaluator.AtifEvalSampleList

Build ATIF-native samples for all eval input items.