StarWhisper Project Overview
The StarWhisper project aims to push the boundaries of astronomical research through advanced AI modeling. Developed with support from the National Astronomical Observatory of China and the Zhijiang Laboratory, the StarWhisper series includes language models, time sequence models, and multimodal models with parameters ranging from 7 billion to 72 billion.
Key Updates
-
Data Refinement and Training Advances: The project has enhanced its model's performance in astrophysics, coding, and Agent capabilities by cleaning and correcting a feedback loop of popular scientific and research data.
-
StarWhisper LC Report: A technical report has been published describing a state-of-the-art large model-based method for astronomical light curve data processing.
-
StarWhisper Pulsar: An upcoming report will introduce a cutting-edge method for pulsar detection using large models.
-
Multimodal Framework and Telescope Integration: The project employs a Visual Agent to establish a multimodal, multitask framework and connect with telescope control systems.
Demonstration of Capabilities
StarWhisper's potential is illustrated through several images that exhibit its astronomical analysis and image processing capabilities.
Quick Start with StarWhisper
To use the StarWhisper model for engaging in interactive dialogues, follow the Python code example provided in the project's documentation. This guide walks you through setting up the model, processing images, and generating descriptive text outputs in interaction scenarios.
The SiTian Project
The SiTian project, as proposed by Chinese astronomers, aims to create a major astronomical infrastructure focused on time-domain astronomy. This initiative plans to install 54 large-field telescopes across selected sites in China to form a multi-band monitoring network. The project is set to conduct high-precision, tri-color surveys of 10,000 square degrees every 30 minutes. The data processing "brain" of the SiTian project requires adept AI tools, and StarWhisper is being explored as a potential solution for integrating astronomical knowledge and addressing specific challenges with multimodal approaches.
Licensing Information
The StarWhisper project is open-source under the Apache-2.0 license. However, the use of specific model weights, such as those for Qwen1.5-14B Chat, must comply with their respective licenses.
Future Goals
Large Language Model (Educational Approach)
- Continue pre-training on additional materials to expand astronomical knowledge.
- Adjust the balance of general and specialized data in supervised fine-tuning to combat catastrophic forgetting.
- Use reinforcement learning from human feedback to enhance model performance.
- Fine-tune specific datasets to improve the model's summarization ability and adaptiveness to knowledge repositories.
- Complete the SiTian-variable star knowledge graph to reduce hallucination issues in the variable star domain.
Specialized Multimodal (Research Tool)
- Release multimodal fine-tuning weights for public use.
- Explore applications of multimodal models in astronomical image generation and recognition.
Observation Agent (SiTian Brain)
- Enhance programming capabilities within the astronomy field.
- Conduct exploratory projects with MiniSiTian/SiTian prototypes for interaction with astronomical environments.
- Consider tool learning to connect with specialized astronomical tools.
- Investigate the feasibility of StarWhisper as a backup solution for the SiTian brain.
Citation Information
For those utilizing the work conducted by the StarWhisper project, please refer to the project's technical paper using the provided citation format.
StarWhisper's Growth
The project has garnered significant attention on GitHub, as illustrated by its rise in star ratings over time, reflecting its impact and community interest.