Windmi1l/geoagenticrag-reranker-v5

GeoAgenticRag Description Reranker v5

Final merged BGE reranker v2 m3 weights fine-tuned to rank candidate description-tree nodes for a query.

File

• model.safetensors — final merged model weights (about 1.14 GB).

This repository contains the final merged weights only. It does not include intermediate training checkpoints or adapters. The checkpoint was selected at epoch 4 using question-disjoint validation weak-label nDCG@10. Labels were generated by rules from production branch queries; they are not expert gold labels.

Base model: BAAI/bge-reranker-v2-m3. License: Apache-2.0.

The matching model configuration and tokenizer are not included in this minimal weights repository. Use the base model files and GeoAgenticRag configuration when loading these weights.

Evaluation snapshot

On an exploratory 100-question MapQA Illinois sample (50 MapQA-ADJ and 50 MapQA-AME), the paired trained-reranker configuration completed 95 questions and scored Recall@1/5/10 of 65.53/75.79/76.32%, MRR 73.32%, NDCG@10 73.66%, and Hit@10 77.89%. Metrics use completed questions only; five cases failed. This configuration pairs the trained reranker with the base JEV. The sample is a component study, not the four-dataset benchmark result.