Model Summary:
Llama-3.1-Centaur-70B is a foundation model of cognition model that can predict and simulate human behavior in any behavioral experiment expressed in natural language.
- Paper: A foundation model to predict and capture human cognition
- More Information: Documentation
- Point of Contact: Marcel Binz
Usage:
Note that Centaur is trained on a data set in which human choices are encapsulated by "<<" and ">>" tokens. For optimal performance, it is recommended to adjust prompts accordingly.
You can use the model using HuggingFace Transformers library with 2 or more 80GB GPUs (NVIDIA Ampere or newer) with at least 150GB of free disk space to accommodate the download.
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "marcelbinz/Llama-3.1-Centaur-70B"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
More details are provided in this example script. You can alternatively run the model with unsloth on a single 80GB GPU using the low-rank adapter.
Licensing Information
Llama 3.1 Community License Agreement
Citation Information
@misc{binz2024centaurfoundationmodelhuman,
title={Centaur: a foundation model of human cognition},
author={Marcel Binz and Elif Akata and Matthias Bethge and Franziska Brändle and Fred Callaway and Julian Coda-Forno and Peter Dayan and Can Demircan and Maria K. Eckstein and Noémi Éltető and Thomas L. Griffiths and Susanne Haridi and Akshay K. Jagadish and Li Ji-An and Alexander Kipnis and Sreejan Kumar and Tobias Ludwig and Marvin Mathony and Marcelo Mattar and Alireza Modirshanechi and Surabhi S. Nath and Joshua C. Peterson and Milena Rmus and Evan M. Russek and Tankred Saanum and Natalia Scharfenberg and Johannes A. Schubert and Luca M. Schulze Buschoff and Nishad Singhi and Xin Sui and Mirko Thalmann and Fabian Theis and Vuong Truong and Vishaal Udandarao and Konstantinos Voudouris and Robert Wilson and Kristin Witte and Shuchen Wu and Dirk Wulff and Huadong Xiong and Eric Schulz},
year={2024},
eprint={2410.20268},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2410.20268},
}