Beyond Scale: the Diversity Coefficient as a Data Quality Metric for Natural Language Datasets

This repository provides the official implementation of the Task2Vec Diversity Coefficient for computing natural language data diversity from the following paper:

Beyond Scale: the Diversity Coefficient as a Data Quality Metric Demonstrates LLMs are Pre-trained on Formally Diverse Data. Alycia Lee, Brando Miranda, Sanmi Koyejo. Paper: https://arxiv.org/abs/2306.13840

This repository also contains code for generating GINC datasets and computing the Diversity Coefficient of those datasets (see ginc/).

Getting Started

diversity/ contains the Task2Vec diversity coefficient computation for natural language data. See #Quick-start for a tutorial of computing the diversity coefficient for a language dataset. Run diversity/runner.sh to compute Task2Vec embeddings and diversity coefficient for c4, WikiText-103, and The Pile.

ginc/ contains Generative In-Context learning Dataset from the original GINC repo.

Run ginc/runner_generate.sh to generate GINC datasets with varying number of HMMs and number of symbols. Run ginc/runner_train.sh to train GPT-2 Transformers on GINC datasets using wandb.

Acknowledgements

We acknowledge that code in ginc/ was sourced from the original GINC repo. We thank Rylan Schaeffer for his contributions to updating the scripts in ginc/ for ease of usage.

Citation

If you found this repo useful, please cite

@article{lee2023scale,
      author={Alycia Lee and Brando Miranda and Sanmi Koyejo},
      journal={arXiv preprint arXiv:2306.13840},
      title={Beyond Scale: the Diversity Coefficient as a Data Quality Metric Demonstrates LLMs are Pre-trained on Formally Diverse Data}, 
      year={2023},
}