llama-nemotron-rerank-1b-v2 | NVIDIA NeMo AutoModel

NeMo AutoModel provides a retrieval variant of Meta’s Llama that defaults to bidirectional attention for reranking. This lets the query and document interact across the full sequence before a classification head produces a relevance score. Set model.is_causal: true to use causal attention. When the option is omitted, a saved text-config value takes precedence over the bidirectional default.

For the bi-encoder variant, see Llama (Bidirectional) for Embedding.

Set up NeMo AutoModel with the latest container or follow the installation instructions.

Train a Reranker with llama-nemotron-rerank-1b-v2

From the repository root, run:

uv run automodel examples/retrieval/cross_encoder/llama3_2_1b.yaml --nproc-per-node 8

Choose a Workflow

Goal	Start Here
Train a Llama 3.2 1B cross-encoder	Use llama3_2_1b.yaml.

Model Reference

Model Architecture

Property	Value
Tasks	Reranking
Architecture	`LlamaBidirectionalForSequenceClassification`
Parameters	1B
Hugging Face Organization	meta-llama

Reranking Models

The cross-encoder path is used for pairwise relevance scoring and reranking.

Architecture	Task	Wrapper Class	Description
`LlamaBidirectionalForSequenceClassification`	Reranking	`NeMoAutoModelCrossEncoder`	Bidirectional Llama with classification head for relevance scoring

Available Models

Model	HF ID
Llama 3.2 1B	`meta-llama/Llama-3.2-1B`
llama-nemotron-rerank-1b-v2	`nvidia/llama-nemotron-rerank-1b-v2`

NVIDIA trained and released the Llama Nemotron Reranking 1B model, optimized to produce a relevance logit score indicating how well a document matches a given query. The model was fine-tuned with a bidirectional attention mechanism for multilingual and cross-lingual question-answer retrieval, with support for long documents (up to 8,192 tokens).