Llama (Bidirectional) for Reranking#
NeMo AutoModel provides a bidirectional variant of Meta’s Llama for reranking tasks. Unlike the standard causal (left-to-right) Llama used for text generation, this variant uses bidirectional attention, allowing the query and document to interact across the full sequence before a classification head produces a relevance score.
For the bi-encoder variant, see Llama (Bidirectional) for Embedding.
Tasks |
Reranking |
Architecture |
|
Parameters |
1B – 8B |
HF Org |
Available Models#
Any Llama checkpoint can be loaded as a bidirectional reranking backbone. The following configurations have been tested:
Llama 3.2 1B — fast iteration, fits on a single GPU
Llama 3.1 8B — higher-quality reranking for production use
Reranking Models#
The cross-encoder path is used for pairwise relevance scoring and reranking.
Architecture |
Task |
Wrapper Class |
Description |
|---|---|---|---|
|
Reranking |
|
Bidirectional Llama with classification head for relevance scoring |
Example HF Models#
Model |
HF ID |
|---|---|
Llama 3.2 1B |
|
Llama 3.1 8B |
Example Recipes#
Recipe |
Description |
|---|---|
Cross-encoder — Llama 3.2 1B reranker |
Try with NeMo AutoModel#
1. Install NeMo AutoModel. Refer to the (Installation Guide) for information:
uv pip install nemo-automodel
2. Clone the repo to get the example recipes:
git clone https://github.com/NVIDIA-NeMo/Automodel.git
cd Automodel
3. Run the recipe from inside the repo:
torchrun --nproc-per-node=8 examples/retrieval/cross_encoder/finetune.py --config examples/retrieval/cross_encoder/llama3_2_1b.yaml
Run with Docker
1. Pull the container and mount a checkpoint directory:
docker run --gpus all -it --rm \
--shm-size=8g \
-v $(pwd)/checkpoints:/opt/Automodel/checkpoints \
nvcr.io/nvidia/nemo-automodel:26.04.00
2. Navigate to the AutoModel directory (where the recipes are):
cd /opt/Automodel
3. Run the recipe:
torchrun --nproc-per-node=8 examples/retrieval/cross_encoder/finetune.py --config examples/retrieval/cross_encoder/llama3_2_1b.yaml
See the Installation Guide.
Hugging Face Model Cards#
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).