nemo_automodel.components.models.d1.records

Published D1 text prompts with explicit question and option identities.

Module Contents

Classes

Name	Description
`D1TextExample`	One decision in the publisher’s internal option order.

Functions

Name	Description
`encode_d1_record`	Encode unchanged SOC records using reviewed pinned publisher prompt code.

API

class nemo_automodel.components.models.d1.records.D1TextExample(
    record_id: str,
    input_ids: tuple[int, ...],
    question_id: str,
    option_ids: tuple[str, ...],
    marker_positions: tuple[int, ...],
    alias_token_ids: tuple[tuple[int, ...], ...],
    head_question_type: int,
    temperature: float
)

Dataclass

One decision in the publisher’s internal option order.

Marker positions and alias groups are mutually exclusive model readouts. Temperature applies to Omni text logits before training and evaluation. Recipes align labels and rewards separately using the question and option IDs.

alias_token_ids

tuple[tuple[int, ...], ...]

head_question_type

int

input_ids

tuple[int, ...]

marker_positions

tuple[int, ...]

option_ids

tuple[str, ...]

question_id

str

record_id

str

temperature

float

nemo_automodel.components.models.d1.records.encode_d1_record(
    record: dict[str, typing.Any],
    tokenizer: typing.Any,
    prompt: types.ModuleType,
    omni: bool,
    max_length: int,
    temperatures: dict[str, float] | None = None
) -> list[nemo_automodel.components.models.d1.records.D1TextExample]

Encode unchanged SOC records using reviewed pinned publisher prompt code.

Parameters:

record

dict[str, Any]

Validated state/questions record. Supervision is not consumed.

tokenizer

Any

Tokenizer from the same immutable release as the prompt code.

prompt

ModuleType

Imported publisher prompt module, pinned by the recipe.

omni

bool

Select the marker encoder instead of the causal token-alias prompt.

max_length

int

Context limit; causal inputs are rejected rather than truncated.

temperatures

dict[str, float] | NoneDefaults to None

Released Omni text calibration, unchanged across stages.

Returns: list[D1TextExample]

Encoded questions with their model-output option identities. The caller