AstroSage-Llama-3.1-8B

https://arxiv.org/abs/2411.09012

AstroSage-Llama-3.1-8B is a domain-specialized natural-language AI assistant tailored for research in astronomy, astrophysics, and cosmology. Trained on the complete collection of astronomy-related arXiv papers from 2007-2024 along with millions of synthetically-generated question-answer pairs and other astronomical literature, AstroSage-Llama-3.1-8B demonstrates excellent proficiency on a wide range of questions. This achievement demonstrates the potential of domain specialization in AI, suggesting that focused training can yield capabilities exceeding those of much larger, general-purpose models.

Model Details

• Base Architecture: Meta-Llama-3.1-8B
• Base Model: Meta-Llama-3.1-8B
• Parameters: 8 billion
• Training Focus: Astronomy, Astrophysics, Cosmology, and Astronomical Instrumentation
• License: Llama 3.1 Community License
• Development Process:
  1. Continued Pre-training (CPT) on astronomical literature
  2. Supervised Fine-tuning (SFT) on QA pairs and instruction sets
  3. Model merging with Meta-Llama-3.1-8B-Instruct (75% CPT+SFT / 25% Meta-Instruct)

Using the model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model and tokenizer
model = AutoModelForCausalLM.from_pretrained("AstroMLab/AstroSage-8b", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("AstroMLab/AstroSage-8b")

# Function to generate a response
def generate_response(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    outputs = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )
    response = outputs[0][inputs['input_ids'].shape[-1]:]
    decoded = tokenizer.decode(response, skip_special_tokens=True)

    return decoded

# Example usage
prompt = """
You are an expert in general astrophysics. Your task is to answer the following question:
What are the main components of a galaxy?
"""
response = generate_response(prompt)
print(response)

Model Improvements and Performance

AstroSage-Llama-3.1-8B shows remarkable performance improvements:

The model demonstrates:

• Outperformance of all 8B parameter models
• Comparable performance to GPT-4o (80.4%)
• ~1000x more cost-effective than proprietary models
• 7 percentage-point improvement over base Llama-3.1-8b model

Training Data

• Continued Pre-training:

  • ~250,000 arXiv preprints (2007-2024) from astro-ph and gr-qc
  • Astronomy-related Wikipedia articles
  • Selected astronomy textbooks
  • Total: 3.3 billion tokens, 19.9 GB plaintext

• Supervised Fine-tuning:

  • 8.8 million curated QA pairs
  • Filtered Infinity-Instruct-7M dataset
  • Paper summaries and metadata
  • Total: 2.0 billion tokens, 9.8 GB plaintext

Intended Use

• Curiosity-driven question answering
• Brainstorming new ideas
• Astronomical research assistance
• Educational support in astronomy
• Literature review and summarization
• Scientific explanation of concepts

Limitations

• Training data cutoff: January 2024
• As with all LLMs, hallucinations are possible
• Limited by 8B parameter size for complex reasoning
• Paper metadata not perfectly memorized
• Performance primarily validated on multiple-choice questions
• Primarily trained for use in English

Technical Specifications

• Architecture: Based on Meta-Llama 3.1
• Training Infrastructure: ORNL OLCF Frontier
• Hosting: Hugging Face Hub (AstroMLab/AstroSage-8B)

Ethical Considerations

While this model is designed for scientific use:

• Should not be used as sole source for critical research decisions
• Output should be verified against primary sources
• May reflect biases present in astronomical literature

Citation and Contact

• Corresponding author: Tijmen de Haan (tijmen dot dehaan at gmail dot com)
• AstroMLab: astromachinelearninglab at gmail dot com
• Please cite the AstroMLab 3 paper when referencing this model:

@preprint{dehaan2024astromlab3,
      title={AstroMLab 3: Achieving GPT-4o Level Performance in Astronomy with a Specialized 8B-Parameter Large Language Model}, 
      author={Tijmen de Haan and Yuan-Sen Ting and Tirthankar Ghosal and Tuan Dung Nguyen and Alberto Accomazzi and Azton Wells and Nesar Ramachandra and Rui Pan and Zechang Sun},
      year={2024},
      eprint={2411.09012},
      archivePrefix={arXiv},
      primaryClass={astro-ph.IM},
      url={https://arxiv.org/abs/2411.09012}, 
}