TART: A plug-and-play Transformer module for task-agnostic reasoning

TART is a package for training and deploying task-agnostic reasoning modules which can be composed with arbitrary foundation models to improve in-context learning peformance for classification tasks. TART is task, model and domain agnostic.

We introduce TART in our paper, TART: A plug-and-play Transformer module for task-agnostic reasoning.

Note: We are thankful to Garg et al., for their wonderful repository, and acknowledge that we have adopted aspects of their code base.

Getting started

You can start by cloning our repository and following the steps below.

Install the dependencies for our package using Conda.

conda env create -f environment.yml
conda activate tart

Download pre-trained TART Reasoning module (you will need this if you want to run the sample notebooks in src/notebooks)
```
wget https://github.com/HazyResearch/TART/releases/download/initial_release/tart_heads.zip
unzip tart_heads.zip
```
[Optional] If you plan to train TART reasoning modules, populate conf/wandb.yaml with you wandb info.

Training a TART reasoning module

All files needed for training a reasoning module can be found in src/reasoning_module. To train a reasoning module, use the following command:
```
python src/reasoning_module/train.py --config src/reasoning_module/conf/tart_heads/reasoning_head_s258.yaml
```
To customize the specifications of the TART Head (i.e., sequence length, training noise etc.), modify parameters in the config file: reasoning_head_s258.yaml.

Playing with TART!

Sample notebooks for deploying TART on image, text, and audio datasets can be found in src/notebooks. We provide more in-depth walk-throughs on the training process of the TART reasoning module and the inference time procedure of TART in src/notebooks/tart-101.
We encourage first-time users to get familiar with TART via these sample notebooks!

Evaluating TART

All utilities needed to evaluate TART can be found in src/eval. In src/eval/run_scripts, we provide several bash scripts demonstrating how to evaluate TART on audio, image, and text tasks. We refer users to those scripts for more details. We provide our datasets (train / test splits) here.

Maintainers

Avanika Narayan (contact: avanika@cs.stanford.edu)
Kush Bhatia (contact: kushb@cs.stanford.edu)

kgourgou / TART