Skip to navigation
ReferenceFull Library ReferenceNemo AutomodelNemo AutomodelComponentsModelsMinistral Bidirectionalnemo_automodel.components.models.ministral_bidirectional.reranker_model

nemo_automodel.components.models.ministral_bidirectional.reranker_model

View as MarkdownOpen in Claude

Portable Transformers implementation of pooled Mistral3 sequence scoring.

Module Contents

Classes

NameDescription
Mistral3ForSequenceClassificationMistral3 backbone with masked pooling and an FP32 scoring projection.

API

class nemo_automodel.components.models.ministral_bidirectional.reranker_model.Mistral3ForSequenceClassification(
config: transformers.Mistral3Config
)

Bases: Mistral3PreTrainedModel

Mistral3 backbone with masked pooling and an FP32 scoring projection.

base_model_prefix
= 'model'
model
= Mistral3Model(config)
score
nemo_automodel.components.models.ministral_bidirectional.reranker_model.Mistral3ForSequenceClassification.forward(
input_ids: torch.Tensor | None = None,
attention_mask: torch.Tensor | None = None,
pixel_values: torch.Tensor | None = None,
image_sizes: torch.Tensor | None = None,
position_ids: torch.Tensor | None = None,
kwargs: typing.Any = {}
) -> transformers.modeling_outputs.SequenceClassifierOutputWithPast

Score tokenized text or multimodal pairs.

Parameters:

input_ids
torch.Tensor | NoneDefaults to None

Integer tensor of shape [batch, sequence].

attention_mask
torch.Tensor | NoneDefaults to None

Padding mask of shape [batch, sequence]; required for pooling.

pixel_values
torch.Tensor | NoneDefaults to None

Optional image tensor of shape [images, channels, height, width].

image_sizes
torch.Tensor | NoneDefaults to None

Optional integer tensor of shape [images, 2], height then width.

position_ids
torch.Tensor | NoneDefaults to None

Optional integer tensor of shape [batch, sequence].

**kwargs
AnyDefaults to {}

Additional Mistral3Model inputs, following its forward tensor contract.

Returns: SequenceClassifierOutputWithPast

Classifier output with raw, unscaled FP32 logits of shape [batch, num_labels].

classmethod

Load the shared classifier head without changing global Transformers conversion mappings.