common.dataset

Dataset vocabulary shared across the agent packages.

Dependency-light on purpose: the shared layer (common) must stay importable without the ML stack, so the dataset types it defines live here, not in ml.dataset.

Classes

Result(coverage_percent, tests_number, ...)

A map's validation result, ordered so that a better result is greater.

TrainingDatasetMode(value)

Which split of the training dataset is active.

class common.dataset.Result(coverage_percent, tests_number, steps_number, errors_number)[source]

Bases: object

A map’s validation result, ordered so that a better result is greater.

a < b means b is better: higher coverage, then fewer tests, then fewer steps, then more errors. The ordering lives solely in sort_key and __lt__; construction sites pass the plain (non-negated) counts.

The on-disk representation (to_tuple()/from_tuple()) keeps the historical sign-flipped tuple format (coverage, -tests, -steps, errors) so existing processed datasets remain readable.

Parameters:
  • coverage_percent (float)

  • tests_number (int)

  • steps_number (int)

  • errors_number (int)

coverage_percent: float
errors_number: int
classmethod from_tuple(values)[source]

Build a Result from its on-disk (sign-flipped) tuple.

Parameters:

values (Sequence[float | int])

Return type:

Result

property sort_key: Tuple[float, int, int, int]

The lexicographic key implementing the ordering above (larger = better).

steps_number: int
tests_number: int
to_tuple()[source]

Return the on-disk representation: tests and steps negated.

Return type:

Tuple[float, int, int, int]

class common.dataset.TrainingDatasetMode(value)[source]

Bases: Enum

Which split of the training dataset is active.

The member values are serialized into configs and MLflow artifacts, so they are pinned behaviour, not implementation detail.

TRAINING = 'TRAINING'
VALIDATION = 'VALIDATION'