StarWhisper is a series of large language models (LLMs) tailored for astronomy, offering language, time-series, and multimodal capabilities ranging from 7B to 72B parameters. Developed with support from the National Astronomical Observatories and Zhijiang Laboratory, it aims to serve as an AI tool for astronomical data processing, particularly for projects like the SITIAN survey, by integrating astronomical knowledge and exploring multimodal solutions for specific challenges.

How It Works

The project leverages a data flywheel approach, refining training methods with cleaned and corrected scientific and popular science data to enhance astronomical physics, coding, and agent capabilities. It has released technical reports on specialized models: StarWhisper Pulsar for state-of-the-art pulsar identification using multimodal LLMs, StarWhisper LC for light curve classification via transfer learning and LLMs, and StarWhisper Telescope for telescope control workflows using LLM agents, which has been applied to the SITIAN project.

Quick Start & Requirements

• Weights for StarWhisper 4.0 are planned for release on the "ModelScope" platform.
• The training dataset for StarWhisper 3 is available in the LLM_Data directory.
• Code related to published papers is available for testing.

Highlighted Details

• StarWhisper Telescope is an agent-based observation assistant system designed to function as an AI astrophysicist.
• The project has developed a multimodal LLM for pulsar identification and a method for light curve classification.
• It aims to improve astronomical agent capabilities and integrate with astronomical professional tools.

Maintenance & Community

• The project has multiple authors and has published several technical reports and a paper on arXiv.
• Further details on community engagement or roadmaps are not explicitly provided in the README.

Licensing & Compatibility

• Project source code is licensed under Apache-2.0.
• Model weights for Qwen Chat require adherence to their respective licenses.

Limitations & Caveats

The project is actively under development, with a to-do list including further fine-tuning for scientific data, reinforcement learning from human feedback, and the development of an astronomical knowledge graph to mitigate hallucinations. Open-sourcing of multimodal fine-tuning weights is pending.