AstroLLaMA-3-8B-Base_AIC

AstroLLaMA-3-8B is a specialized base language model for astronomy, developed by fine-tuning Meta's LLaMA-3-8b architecture on astronomical literature. This model was developed by the AstroMLab team. It is designed for next token prediction tasks and is not an instruct/chat model.

Model Details

• Base Architecture: LLaMA-3-8b
• Training Data: Abstract, Introduction, and Conclusion (AIC) sections from arXiv's astro-ph category papers
• Data Processing: Optical character recognition (OCR) on PDF files using the Nougat tool, followed by summarization using Qwen-2-8B and LLaMA-3.1-8B.
• Fine-tuning Method: Continual Pre-Training (CPT) using the LMFlow framework
• Training Details:
  • Learning rate: 2 × 10⁻⁵
  • Total batch size: 96
  • Maximum token length: 512
  • Warmup ratio: 0.03
  • No gradient accumulation
  • BF16 format
  • Cosine decay schedule for learning rate reduction
  • Training duration: 1 epoch
• Primary Use: Next token prediction for astronomy-related text generation and analysis
• Reference: Pan et al. 2024

Generating text from a prompt

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("AstroMLab/astrollama-3-8b-base_aic")
model = AutoModelForCausalLM.from_pretrained("AstroMLab/astrollama-3-8b-base_aic", device_map="auto")

# Create the pipeline with explicit truncation
from transformers import pipeline
generator = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    truncation=True,
    max_length=512
)

# Example prompt from an astronomy paper
prompt = "In this letter, we report the discovery of the highest redshift, " \
    "heavily obscured, radio-loud QSO candidate selected using JWST NIRCam/MIRI, " \
    "mid-IR, sub-mm, and radio imaging in the COSMOS-Web field. "

# Set seed for reproducibility
torch.manual_seed(42)

# Generate text
generated_text = generator(prompt, do_sample=True)
print(generated_text[0]['generated_text'])

Model Limitations and Biases

A key limitation identified during the development of this model is that training solely on astro-ph data may not be sufficient to significantly improve performance over the base model, especially for the already highly performant LLaMA-3 series. This suggests that to achieve substantial gains, future iterations may need to incorporate a broader range of high-quality astronomical data beyond arXiv, such as textbooks, Wikipedia, and curated summaries.

Here's a performance comparison chart based upon the astronomical benchmarking Q&A as described in Ting et al. 2024:

Model	Score (%)
AstroSage-LLaMA-3.1-8B (AstroMLab)	80.9
LLaMA-3.1-8B	73.7
LLaMA-3-8B	72.9
AstroLLaMA-3-8B-Base_AIC (AstroMLab)	71.9
Gemma-2-9B	71.5
Qwen-2.5-7B	70.4
Yi-1.5-9B	68.4
InternLM-2.5-7B	64.5
Mistral-7B-v0.3	63.9
ChatGLM3-6B	50.4
AstroLLaMA-2-7B-AIC	44.3
AstroLLaMA-2-7B-Abstract	43.5

As shown, while AstroLLaMA-3-8B performs competitively among models in its class, it does not surpass the performance of the base LLaMA-3-8B model. This underscores the challenges in developing specialized models and the need for more diverse and comprehensive training data.

This model is released primarily for reproducibility purposes, allowing researchers to track the development process and compare different iterations of AstroLLaMA models.

For optimal performance and the most up-to-date capabilities in astronomy-related tasks, we recommend using AstroSage-8B, where these limitations have been addressed. The newer model incorporates expanded training data beyond astro-ph and features a greatly expanded fine-tuning process, resulting in significantly improved performance.

Ethical Considerations

While this model is designed for scientific use, users should be mindful of potential misuse, such as generating misleading scientific content. Always verify model outputs against peer-reviewed sources for critical applications.

Citation

If you use this model in your research, please cite:

@ARTICLE{2024arXiv240919750P,
       author = {{Pan}, Rui and {Dung Nguyen}, Tuan and {Arora}, Hardik and {Accomazzi}, Alberto and {Ghosal}, Tirthankar and {Ting}, Yuan-Sen},
        title = "{AstroMLab 2: AstroLLaMA-2-70B Model and Benchmarking Specialised LLMs for Astronomy}",
      journal = {arXiv e-prints},
     keywords = {Astrophysics - Instrumentation and Methods for Astrophysics, Computer Science - Computation and Language},
         year = 2024,
        month = sep,
          eid = {arXiv:2409.19750},
        pages = {arXiv:2409.19750},
          doi = {10.48550/arXiv.2409.19750},
archivePrefix = {arXiv},
       eprint = {2409.19750},
 primaryClass = {astro-ph.IM},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2024arXiv240919750P},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}