Official repo for paper LeTI: Learning to Generate from Textual Interactions.
This repo contains code that can be used to reproduce the experiments in the paper. Train/evaluation code are written using Jax/Flax to train on Google Cloud's TPU VM instances. Research is supported with Cloud TPUs from Google's TPU Research Cloud (TRC).
WARNING Training and evaluation of LeTI requires executing untrusted model-generated code. Users are strongly encouraged to sandbox the code execution so that it does not perform destructive actions on their host or network.
You can setup your Google Cloud TPU and Storage following docs/SETUP.md. Alternatively, you may also adapt the released code to your specific computing setup.
You can prepare datasets for training and evaluation following instructions in docs/DATA.md.
Since the training and evaluation code is implemented using Jax/Flax, you will need to convert huggingface model checkpoints (pytorch) into T5X format, following instructions in docs/MODEL.md. We release the 350M and 2B model checkpoints here.
You can follow docs/TRAIN.md and docs/EVAL.md to train or evaluate a specific model.
@article{Wang2023LeTI,
title={LeTI: Learning to Generate from Textual Interactions},
author={Xingyao Wang and Hao Peng and Reyhaneh Jabbarvand and Heng Ji},
journal={ArXiv},
year={2023},
volume={abs/2305.10314},
}