🍮 The WHOLE FLAN Collection! 🍮

Overview

This repository includes the full dataset from the FLAN Collection, totalling ~300GB as parquets. Generated using the official seqio templating from the Google FLAN Collection GitHub repo. The data is subject to all the same licensing of the component datasets.

To keep up with our continued work on OpenOrca and other exciting research, find our Discord here:

https://AlignmentLab.ai

Motivation

This work was done as part of the requirements for the OpenOrca project. There was not a large enough subset of FLAN Collection generated publicly to subsample from to complete the work. So, we opted to process the entire collection ourselves. Generating this requires an understanding of seqio and a Linux server with 512GB of CPU ram, as well as fast drives and custom limits for many parameters beyond what is default on Linux server distributions (e.g., requiring up to 45,000 threads running at once). It takes downloading over 400GB of datasets, working around tfds bugs, and then processing the datasets over the course of several days. We provide this repo as a resource to other ML researchers, as it saves these time consuming and laborious steps to getting the data into a more accessible format for further consumption.

Data

Organization

• JSON files at top level are used for subsampling in OpenOrca
• Parquets in subdirectories contain the entire FLAN collection in Dask-sharded folders by submix fractions

Zero-Shot vs Few-Shot and Options vs No-Options

The core sub-collections of FLAN are CoT, Dialog, NIv2, T0, and flan2021. Within those sub-collections are four "remixes" of the data that are templated differently:

• Zero-Shot and Few-Shot
  • Zero-Shot provides a prompt, question, or challenge without any exemplaries prior
  • Few-Shot provides exemplaries first
• Options and No-Options
  • Options provides a question or challenge with multiple-choice (e.g. A/B/C/D) answer options provided to select from
  • No-Options requires a free-form answer

For every sub-collection, only some of the "remixes" may officially be provided. All available have been generated in full without any redaction or sub-sampling.

An example: t0_fsopt_data folder contains the sub-collection T0's Few-Shot (FS), Options (OPT) remix set. Notably, this is the largest "remix" and the one that necessitates 512GB CPU ram to generate. The raw json output is nearly 200GB.

Parquet Sizes

Each sub-collection's individual remixes are provided as Parquet files which have been sharded by Dask into ~160MB chunks (starting from 256MB blocks of the source jsonl files). The folder structure along with size sums is provided below.

$ du -h --max-depth=1 ./
9.1G    ./niv2_fsopt_data
2.4G    ./niv2_zsopt_data
59G     ./flan_fsopt_data
984M    ./dialog_zsopt_data
11G     ./flan_zsopt_data
8.6G    ./dialog_fsopt_data
16G     ./t0_zsnoopt_data
149M    ./cot_fsopt_data
20M     ./cot_zsopt_data
17G     ./t0_zsopt_data
11G     ./flan_zsnoopt_data
101G    ./t0_fsopt_data
25G     ./flan_fsnoopt_data
39G     ./t0_fsnoopt_data
296G    ./

Citations

@misc{goodson2023huggyflan
      title={Fine FLAN: Seqio to Parquet So You Don't Have To},
      author={Bleys Goodson},
      year={2023},
      publisher = {HuggingFace},
      journal = {HuggingFace repository},
      howpublished = {\url{https://https://huggingface.co./datasets/Open-Orca/FLAN},
}

@misc{longpre2023flan,
      title={The Flan Collection: Designing Data and Methods for Effective Instruction Tuning}, 
      author={Shayne Longpre and Le Hou and Tu Vu and Albert Webson and Hyung Won Chung and Yi Tay and Denny Zhou and Quoc V. Le and Barret Zoph and Jason Wei and Adam Roberts},
      year={2023},
      eprint={2301.13688},
      archivePrefix={arXiv},
      primaryClass={cs.AI}
}

@misc{wei2022finetuned,
      title={Finetuned Language Models Are Zero-Shot Learners}, 
      author={Jason Wei and Maarten Bosma and Vincent Y. Zhao and Kelvin Guu and Adams Wei Yu and Brian Lester and Nan Du and Andrew M. Dai and Quoc V. Le},
      year={2022},
      eprint={2109.01652},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@misc{sanh2022multitask,
      title={Multitask Prompted Training Enables Zero-Shot Task Generalization}, 
      author={Victor Sanh and Albert Webson and Colin Raffel and Stephen H. Bach and Lintang Sutawika and Zaid Alyafeai and Antoine Chaffin and Arnaud Stiegler and Teven Le Scao and Arun Raja and Manan Dey and M Saiful Bari and Canwen Xu and Urmish Thakker and Shanya Sharma Sharma and Eliza Szczechla and Taewoon Kim and Gunjan Chhablani and Nihal Nayak and Debajyoti Datta and Jonathan Chang and Mike Tian-Jian Jiang and Han Wang and Matteo Manica and Sheng Shen and Zheng Xin Yong and Harshit Pandey and Rachel Bawden and Thomas Wang and Trishala Neeraj and Jos Rozen and Abheesht Sharma and Andrea Santilli and Thibault Fevry and Jason Alan Fries and Ryan Teehan and Tali Bers and Stella Biderman and Leo Gao and Thomas Wolf and Alexander M. Rush},
      year={2022},
      eprint={2110.08207},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

@misc{wang2022supernaturalinstructions,
      title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks}, 
      author={Yizhong Wang and Swaroop Mishra and Pegah Alipoormolabashi and Yeganeh Kordi and Amirreza Mirzaei and Anjana Arunkumar and Arjun Ashok and Arut Selvan Dhanasekaran and Atharva Naik and David Stap and Eshaan Pathak and Giannis Karamanolakis and Haizhi Gary Lai and Ishan Purohit and Ishani Mondal and Jacob Anderson and Kirby Kuznia and Krima Doshi and Maitreya Patel and Kuntal Kumar Pal and Mehrad Moradshahi and Mihir Parmar and Mirali Purohit and Neeraj Varshney and Phani Rohitha Kaza and Pulkit Verma and Ravsehaj Singh Puri and Rushang Karia and Shailaja Keyur Sampat and Savan Doshi and Siddhartha Mishra and Sujan Reddy and Sumanta Patro and Tanay Dixit and Xudong Shen and Chitta Baral and Yejin Choi and Noah A. Smith and Hannaneh Hajishirzi and Daniel Khashabi},
      year={2022},
      eprint={2204.07705},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}