Nemotron Reranking Fine-Tuning

Adapt a reranker to order the documents returned by a first-stage retriever. Use this workflow when relevant documents reach the candidate set but rank below less relevant results. If candidate recall is low, start with Nemotron Embedding Fine-Tuning.

Choose Your Workflow

The two profiles use different data preparation and deployment paths:

Model	Profile	Input	Start Here
Nemotron 3.5 Rerank 8B	nemotron-3.5	Text-only, image-only, or paired documents prepared by the embedding recipe	Nemotron 3.5 workflow
Llama Nemotron Rerank 1B v2	default	Text	Fine-tune rerankers on text data

For Nemotron 3.5, prepare data with the embedding workflow, then run reranker training, evaluation, and deployment explicitly. Use a separate reranker artifact_root and pass the prepared data paths. The consolidated checkpoint deploys through vLLM without Stage 4 export.

The default Llama workflow includes text generation, preparation, training, evaluation, ONNX/TensorRT export, and NIM deployment.

Plan and Verify a Run

For the text QA workflow, start with the quick start, then inspect evaluation metrics and served-model accuracy.

For the Nemotron 3.5 workflow, use these task guides:

• Plan checkpoint resources before full 8B fine-tuning.

• Prepare multimodal data, retaining the complete source bundle and corpus references.

• Evaluate base and fine-tuned checkpoints on a fixed candidate set.

• Deploy and compare served scores using the saved candidates and raw relevance scores.

• Inspect shared data contracts and configure multimodal generation.