Reranker Models

NeMo 2.0 supports fine-tuning a causal language model into a reranker model. NeMo 2.0 uses NeMo-Run to facilitate scaling across multiple GPUs. Reranker models are specialized neural networks designed to improve search and retrieval systems by re-scoring a set of candidate items (such as documents or passages) based on their relevance to a query. Unlike embedding models that encode queries and documents into vector space, rerankers directly compare query-document pairs to produce relevance scores, often achieving higher precision by capturing complex semantic relationships between the query and potential matches.

The following reranker models are currently supported in NeMo 2.0:

• Llama Reranker

Default configurations are provided for each model and are outlined in the model-specific documentation linked above. Every configuration can be modified to train on new datasets or to test new model hyperparameters.