AstroLLM: A Domain-Specific Large Language Model for Astronomical Research

AstroLLM is an open-source domain-specific large language model system for astronomy and astrophysics research, designed to address the hallucination problem of general-purpose large language models in professional scientific research scenarios. It is deeply integrated with astronomical databases like NASA ADS and SIMBAD through RAG technology, providing retrieval-augmented answers with real citations, and is positioned as an intelligent research assistant for scientists.

In the field of astronomy, general-purpose large models struggle to provide accurate and reliable scientific research assistance, and the hallucination problem is particularly fatal. AstroLLM's design goal is to become a research assistant that can cite real papers and query real databases, and refuse to answer when evidence is insufficient instead of making up information. Compared to existing astronomical models (e.g., AstroSage), its differentiators include: tool integration capabilities (connecting to databases like SIMBAD and NASA ADS), RAG architecture (real-time knowledge updates), educational adaptability (supporting Socratic teaching for users at different levels), and hardware friendliness (the 8B parameter model can run on consumer-grade hardware).

AstroLLM adopts a layered architecture:

Data and Model Layer

It uses QLoRA supervised fine-tuning based on the Qwen3-4B/8B model, with training data from an astronomical literature corpus, and injects domain knowledge via LoRA.

Retrieval and Tool Layer

The RAG system builds vector storage based on PostgreSQL+pgvector. The tool integration layer bridges multiple data sources: NASA ADS (15 million+ papers), SIMBAD (20 million+ celestial objects), NASA Exoplanet Archive (5,800+ planets), NED (extragalactic object data), and VizieR (23,000+ catalogs).

Service Layer

Inference supports deployment via vLLM and llama.cpp, and the web interface uses the TanStack Start+Elysia tech stack.

AstroLLM iterates in phases; currently it is in Phase 0:

Phase	Timeline	Core Deliverables
Phase1(v1)	1-3 months	Retrieval-augmented assistant: QLoRA SFT, RAG+ADS/SIMBAD, beta version launch
Phase2(v2)	4-8 months	Serious astronomical model: Full LoRA8B, DPO training, expanded toolset
Phase3(v3)	9-18 months	Scientific tool ecosystem: Model family (Nano3B+Core8B+Pro32B), continuous learning
Phase4+(v4+)	From Year 2	Multimodal knowledge base: AION-1 visual bridge, spectrum and light curve processing

AstroLLM's application scenarios include:

1. Literature review: Quickly locate relevant research based on ADS and generate review summaries with citations
2. Celestial object query: Use natural language to query SIMBAD for astrophysical parameters
3. Teaching assistance: Adjust the depth of explanations according to user level to support astronomy education
4. Data analysis: Perform basic astronomical calculations and data processing in combination with Astropy

AstroLLM is an open-source project licensed under Apache 2.0, and actively integrates into the astronomical AI ecosystem: it draws on AstroMLab's benchmarking methods, Multimodal Universe's multimodal datasets, and AION-1's multimodal foundation model experience, and encourages wide adoption and contributions from academia and industry.

AstroLLM represents a typical paradigm for domain-specific large models: building a complete system of tool integration, retrieval augmentation, and knowledge updates, rather than simply fine-tuning general-purpose models. For astronomical researchers, a trustworthy AI assistant is moving from concept to reality.