Reranking Models

Introduction

Reranking models use cross-encoders to score a query-document pair jointly. They are typically used after an embedding model has produced an initial candidate set. NeMo AutoModel supports optimized bidirectional Llama rerankers and a causal Qwen3 reranker that scores with a yes or no decision at the final token, and falls back to Hugging Face AutoModelForSequenceClassification for other architectures.

For first-stage dense retrieval, see Embedding Models.

Optimized Backbones

The Llama backbones convert to bidirectional attention. Qwen3RerankerForCausalReranking stays causal: it reuses the pretrained lm_head and scores a pair as logit("yes") - logit("no") at the final non-padding token, adding no parameters. Its checkpoints serialize as a plain Qwen3ForCausalLM, so they load in stock Transformers and vLLM without custom code.

Hugging Face Auto Backbones

Unregistered scoring architectures load through Hugging Face AutoModelForSequenceClassification. An explicit model.is_causal value overrides a saved text-config value. Without either setting, the fallback preserves the backbone’s native attention mode. Verify that an unregistered architecture honors the selected policy before using it for retrieval training.

Composite backbones must identify their text tower through matching get_text_config(decoder=True) and get_decoder() results. Only that tower’s attention policy is changed.

Supported Models

This table combines recipe-backed checkpoints with documented model families. Dates show when the current checkpoint first appeared in a recipe, or when a documentation-only family page was added. See the combined model support log for recipe-backed checkpoints of every model type.

Supported Workflows

• Fine-tuning (Cross-Encoder): Cross-entropy training on query-document pairs to produce rerankers
• LoRA and PEFT: Parameter-efficient fine-tuning for reranking backbones

Dataset

Retrieval fine-tuning requires query-document pairs: each example is a query paired with one positive document and one or more negative documents. Both inline JSONL and corpus ID-based JSON formats are supported. See the Retrieval Dataset guide.

Train Reranking Models

For a complete walkthrough of training configuration, model-specific settings, and launch commands, see the Embedding and Reranking Fine-Tuning Guide.