# On the Origin of LLMs: An Evolutionary Tree and Graph for 15,821 Large Language Models

**Authors:** Sarah Gao, Andrew Kean Gao  
**arXiv:** [2307.09793](https://arxiv.org/abs/2307.09793)  
**Date:** 2023-07-19  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.09793v1/

## Key points

- Large Language Models (LLMs) like ChatGPT and Bard have gained popularity, attracting millions of users
- Hugging Face platform has seen hundreds of new LLMs being announced each week
- Nearly 16,000 Text Generation models have been uploaded to the site
- Researchers developed a method using hierarchical clustering techniques, n-grams, and term frequency-inverse document frequency to identify communities among LLMs
- Constellation web application serves as an atlas of 15,821 LLMs with visualizations like dendrograms, graphs, word clouds, and scatter plots
- Users can explore different LLMs based on preferences and access statistics and interactive scatter plots comparing likes versus downloads for each model
- Louvain method is used to generate a graph highlighting relationships between different LLMs and clusters within the dataset
- Study provides comprehensive understanding of LLM landscape by identifying popular backbones, settings, training methods, and families
- Constellation web application facilitates exploration of insights visually and keeps up with rapid growth of LLMs

## Summary

In recent years, Large Language Models (LLMs) have gained significant popularity, with models like ChatGPT and Bard attracting millions of users. The Hugging Face platform, which serves as a repository for machine learning models and datasets, has seen hundreds of new LLMs being announced each week. To date, nearly 16,000 Text Generation models have been uploaded to the site. However, despite this massive influx of LLMs, there is no comprehensive index available to understand the popular or trending LLM backbones, settings, training methods and families. To address this gap in knowledge, researchers have taken advantage of the systematic nomenclature used for Hugging Face LLMs. They have developed a method that utilizes hierarchical clustering techniques along with n-grams and term frequency-inverse document frequency to identify communities among LLMs. This approach successfully identifies families of LLMs and accurately clusters them into meaningful subgroups. To make their findings accessible to the public, the researchers have created a web application called Constellation. This application serves as an atlas of 15,821 LLMs and offers various visualizations such as dendrograms, graphs, word clouds and scatter plots. Users can navigate through Constellation to explore different LLMs based on their preferences. Additionally, the web application provides useful statistics and an interactive scatter plot that compares likes versus downloads for each model. The researchers also generate a graph of the models using the Louvain method to detect communities among similar models. This graph helps visualize relationships between different LLMs and highlights clusters within the dataset. Overall, this study presents a comprehensive approach to understanding the landscape of LLMs by identifying popular backbones settings training methods and families. The Constellation web application provides a user-friendly interface for exploring these insights visually. With the rapid growth of LLMs in recent years tools like Constellation are crucial in keeping up with these developments and facilitating further research in the field.