Geospatial large language model trained with a simulated environment for generating tool-use chains autonomously
we propose a framework for training a Geospatial large language model to generate Tool-use Chains autonomously (GTChain). Specifically, we design a seed task-guided self-instruct strategy to generate a geospatial tool-use instruction tuning dataset within a simulated environment, which includes a simulated multiple-step processing environment and simulated geospatial data stream. Subsequently, an open-source general-domain LLM, LLaMA-2-7B, is fine-tuned on the collected instruction data to understand geospatial tasks and learn how to generate geospatial tool-use chains. Finally, we also collect an evaluation dataset to serve as a benchmark for assessing the geospatial tool-use ability of LLMs. Based on the evaluation dataset, experimental results demonstrate that the fine-tuned LLM can effectively solve geospatial tasks with the provided tools, which validates the effectiveness of our framework.
We release training data samples for training GTChain, stored in TrainingData.json. You can contact the author to obtain all the training data.
We release an evaluation dataset to serve as a benchmark for assessing the geospatial tool-use ability of LLMs, stored in EvalData.json.
We provide detailed descriptions of the tools we collected. Details of the tool set (27 tools) are stored in Tools.json, and the external tools (5 tools) are stored in ExternalTools.json.
GTChain and all our publicly available data are intended for research preview and non-commercial use only, subject to the model License of LLaMA2. Please contact us if you find any potential violations. If you have any questions, you can emaill us [email protected].