README_EN.md

StarWhisper 4.0

中文   ｜   English  

Under the support of the NAOC and ZheJiang Lab, we developed StarWhisper 4.0, a series of astronomical models including language models, time-series models, and multi-modal models (7B-72B).

---

Version Updates

1. Data & Training Enhancements

   • Improved astronomical physics, coding, and agent capabilities through refined training methods and cleaned scientific/popular science datasets.
   • Open-sourced the StarWhisper 3 training dataset in the LLM_Data directory.
   • StarWhisper 4.0 weights will be released on ModelScope.

2. StarWhisper Pulsar

   • Technical report on a SOTA multimodal large model for pulsar identification.

3. StarWhisper LC

   • Light curve classification method based on transfer learning and large models.
   • Test code related to the paper is uploaded.
   

4. StarWhisper Telescope

   • Agent-based telescope control workflow for the Near-Neighbor Galaxy Survey System (NGSS).
   • Code is open-sourced in the NGSS directory.
   

---

Demonstration

---

Sitian Project

Sitian is a major astronomical infrastructure proposed by Chinese astronomers for time-domain astronomy. Phase I involves deploying 54 wide-field telescopes (1-meter aperture, 18 groups) across multiple observation sites in China. These telescopes will form a multi-band monitoring network, enabling high-precision three-color "gaze" surveys of 10,000 square degrees every 30 minutes. With a sampling frequency two orders of magnitude higher than global peers, Sitian will:

• Discover new celestial objects/phenomena in extreme energy bursts, gravitational wave counterparts, exoplanets, and solar system bodies.
• Address key scientific questions (e.g., dark matter, black holes, cosmic origins) and national space security (e.g., planetary defense).

As the AI core of Sitian's "brain," StarWhisper integrates astronomical knowledge via large models and explores multimodal solutions for domain-specific challenges.

---

License

• Source code: Apache-2.0 License
• Qwen Chat model weights: Subject to their respective licenses.

---

To-Do List

Large Language Model (Science Communication)

• Optimize the ratio of general vs. domain-specific data during SFT to mitigate catastrophic forgetting.
• Improve performance via reinforcement learning with human feedback (RLHF).
• Enhance summarization capabilities through domain-adaptive fine-tuning.
• Build an astronomical knowledge graph to reduce hallucinations.

Multi-Modal Models (Research Tools)

• Release multimodal fine-tuning weights.
• Explore applications in astronomical image generation and recognition.

Observation Agents (Sitian Brain)

• Boost coding proficiency in astronomy.
• Develop agents for interaction with MiniSiTian/Sitian prototypes.
• Integrate astronomical tools (e.g., ASTROLABE, CASA) via tool learning.
• Validate feasibility as a Sitian brain candidate.

---

Citation

@misc{wang2024starwhispertelescopeagentbasedobservation,
      title={StarWhisper Telescope: Agent-Based Observation Assistant System to Approach AI Astrophysicist}, 
      author={Cunshi Wang and Xinjie Hu and Yu Zhang and Xunhao Chen and Pengliang Du and Yiming Mao and Rui Wang and Yuyang Li and Ying Wu and Hang Yang and Yansong Li and Beichuan Wang and Haiyang Mu and Zheng Wang and Jianfeng Tian and Liang Ge and Yongna Mao and Shengming Li and Xiaomeng Lu and Jinhang Zou and Yang Huang and Ningchen Sun and Jie Zheng and Min He and Yu Bai and Junjie Jin and Hong Wu and Chaohui Shang and Jifeng Liu},
      year={2024},
      eprint={2412.06412},
      archivePrefix={arXiv},
      primaryClass={astro-ph.IM},
      url={https://arxiv.org/abs/2412.06412}, 
}

---

Star History