Terra: Qwen3-VL LoRA fine-tuning on RSRCC

We fine-tuned the Qwen/Qwen3-VL-4B-Instruct VLM on the google/RSRCC dataset. Small VLMs are excellent at describing a single image, but they struggle with temporal and bi-temporal remote sensing tasks. When given two aerial images taken at different times ($T_0$ and $T_1$), generic models often fail to capture subtle differences, suffer from hallucinations, or confuse the chronological order of events.

Installation

Ensure you have uv installed.

To install dependencies and prepare the virtual environment, run:

uv sync

How to Run

To run the training pipeline, activate the environment and execute:

uv run src/train.py

To view the training progress and metrics in the MLflow UI, run:

mlflow ui --backend-store-uri sqlite:///mlflow.db

License

Base model (Qwen3-VL 4B) and dataset (google/RSRCC) are licensed under Apache-2.0 license.

This project ("Terra") is currently source available.