149 lines
4.9 KiB
Markdown
149 lines
4.9 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- llama-3.1
|
|
- astronomy
|
|
- astrophysics
|
|
- cosmology
|
|
- arxiv
|
|
inference: false
|
|
base_model:
|
|
- meta-llama/Meta-Llama-3.1-8B
|
|
---
|
|
|
|
# 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
|
|
|
|
```python
|
|
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:
|
|
|
|
| Model | Score (%) |
|
|
|-------|-----------|
|
|
| **<span style="color:green">AstroSage-Llama-3.1-8B</span>** | **<span style="color:green">80.9</span>** |
|
|
| GPT-4o | 80.4 |
|
|
| LLaMA-3.1-8B | 73.7 |
|
|
| 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 |
|
|
|
|
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},
|
|
}
|
|
``` |