This repository contains replication code for the following paper:

Smoothie: Label Free Language Model Routing
Neel Guha*, Mayee Chen*, Trevor Chow, Ishan Khare, Christopher Ré
NeurIPS 2024
paper | blog

Install the dependencies using the following commands:

> conda create -n "smoothie" python=3.10 -y
> conda activate smoothie
> pip install -r requirements.txt

We store all datasets, predictions, and results from the paper in a HuggingFace dataset repository. You can download the dataset from HuggingFace by running the following command:

> huggingface-cli login --token $HUGGINGFACE_TOKEN --add-to-git-credential
> git clone https://huggingface.co/datasets/hazyresearch/smoothie_data

In tutorials/tutorial.ipynb, we walk through how to use the Smoothie algorithm. The tutorial can be easily adapted for your use case given that you provide a .jsonl file with the dataset inputs, and several json files each containing a different model/prompt's generations.

If you are interested in the mathematical derivation of Smoothie, check out tutorials/algorithm.ipynb.

See reproducing_experiments.md for instructions on how to reproduce the experiments in the paper.

The repository contains the following folders:

• dataset_configs: Contains the configuration files for all single-task and multi-task datasets.
• plots: Contains plots for the paper.
• prompt_templates: Contains the prompt templates for all single-task and multi-task datasets.
• replication_scripts: Contains bash scripts for running experiments in the paper.
• src: Contains the source code for formatting datasets, getting generations, running routing methods, and evaluating results. The subfolder paper contains code for producing the tables and plots in the paper.
• tables: Contains latex tables for the paper.
• tutorials: Contains tutorials for using Smoothie.

If you use Smoothie in your work, please cite the following paper:

@misc{guha2024smoothielabelfreelanguage,
      title={Smoothie: Label Free Language Model Routing}, 
      author={Neel Guha and Mayee F. Chen and Trevor Chow and Ishan S. Khare and Christopher Ré},
      year={2024},
      eprint={2412.04692},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2412.04692}, 
}