EarthJev
Weights of EarthJev: Efficient Visual Evidence Acquisition and Decision for UHR Remote Sensing Understanding. EarthJev keeps a multimodal answering model frozen and adds two small trained modules: a region head that scores image regions for the evidence search (EDM) and a pointer verifier that checks the boxes and points stated in a question (SPV).
Code, fitted parameters and documentation: EarthJev on GitHub.
| Path | Content |
|---|---|
Qwen3.5-9B/ |
the answering model: an unmodified copy of Qwen/Qwen3.5-9B, with its license and model card |
region_head/qwen35-9b/, verifier/qwen35-9b/ |
the region head and the pointer verifier for Qwen3.5-9B |
region_head/qwen3-vl-8b/, verifier/qwen3-vl-8b/ |
the same modules for Qwen/Qwen3-VL-8B-Instruct |
Usage
In a clone of the code repository, with the environment variables of its README:
hf download daidaidaidai1dai/EarthJev --local-dir "$EARTHJEV_MODELS"
python scripts/check_env.py --hash-model
python scripts/run_earthjev.py --bench mme_rs
License
The region heads and pointer verifiers are released under the Apache License 2.0 (LICENSE). Qwen3.5-9B is
redistributed unmodified under its own Apache License 2.0 (Qwen3.5-9B/LICENSE).