Windmi1l/geoagenticrag-finetuned-models

GeoAgenticRag fine-tuned models

This repository contains two merged checkpoints from the GeoAgenticRag project. The checkpoints are stored in subdirectories so each can be loaded independently.

JEV

Path: models/jev-merged/

JEV is a Qwen3.5-0.8B model fine-tuned for structured geospatial agent decisions. Training used simulated states and rule-generated labels. The selected merged checkpoint is from epoch 2. The config, tokenizer, chat template, deployment metadata, training provenance, and license are included beside the weights.

Base model: Qwen3.5-0.8B.

Description reranker v5

Path: models/reranker-v5-merged/

This BGE reranker v2 m3 checkpoint scores candidate description-tree nodes for a query. The selected merged checkpoint is from epoch 4, based on question-disjoint validation weak-label nDCG@10. Labels came from production branch queries and were generated by rules; they are not expert gold labels.

Base model: BAAI/bge-reranker-v2-m3.

Loading

Use the model directory as the from_pretrained path with a Transformers version that supports Qwen3.5 for JEV. For reranking, load the reranker directory with AutoModelForSequenceClassification and its tokenizer. Both base licenses are Apache-2.0; each checkpoint directory contains the applicable license text.

Intended use

These are research checkpoints. Review generated decisions and reranking results before using them in an operational system. The validation labels are weak labels, so their reported metrics should not be interpreted as expert-annotated accuracy.