metadata

dataset_info:
  features:
    - name: input
      dtype: string
    - name: output
      dtype: string
    - name: source
      dtype: string
    - name: subset
      dtype: string
  splits:
    - name: train
      num_bytes: 63759065
      num_examples: 23652
    - name: validation
      num_bytes: 6190242
      num_examples: 2042
    - name: test
      num_bytes: 6080212
      num_examples: 2045
  download_size: 45525146
  dataset_size: 76029519
task_categories:
  - text2text-generation
  - text-generation
  - question-answering
  - conversational
  - summarization
  - table-question-answering
language:
  - en
tags:
  - instruction-tuning
pretty_name: longform
size_categories:
  - 10K<n<100K

LongForm

The LongForm dataset is created by leveraging English corpus examples with augmented instructions. We select a diverse set of human-written documents from existing corpora such as C4 and Wikipedia and generate instructions for the given documents via LLMs. Then, we extend these examples with structured corpora examples such as Stack Exchange and WikiHow and task examples such as question answering, email writing, grammar error correction, story/poem generation, and text summarization.

Distribution

The distribution of the LongForm dataset in terms of the source of examples is below. It contains examples generated from raw text corpora via LLMs, structured corpus examples, as well as various NLP task examples such as email writing, grammar error correction, story/poem generation, and text summarization.

Type	Source	Number of Examples
Corpora	C4	10,000
	Wikipedia	5,000
Structured Corpora	Stack Exchange	4,380
	WikiHow	2,500
Tasks	NIv2	3,684
	Big Bench	600
	BEA-GEC	1,203
	Enron	372
Total		27,739

Train		23,652
Validation		2,042
Test		2,045

Models

	All	Recipe Generation	ELI5	Writing Prompts
T0++	10.9	18.7	3.8	10.2
Tk-Instruct	6.3	12.9*	3.6	2.4
Flan-T5	10.6	20.9*	3.5	7.4
Alpaca-LLaMA-7B	14.6	19.5	12.5	11.8
OPT-30B	11.1	18.6	12.2	2.6
LongForm-T5-XL	16.3	20.2	18.3	10.6
LongForm-OPT-2.7B	17.8	15.5	17.9	19.9
LongForm-OPT-6.7B	17.7	16.9	17.2	19.0
LongForm-LLaMA-7B‡	19.7	21.7	18.6	18.9

‡: We can just release the difference between LongForm-LLaMA-7B and pretrained LLaMA-7B publicly due to restrictions of LLaMA models.

Examples

Source	Instruction	Output
C4	What is the likelihood that Donald Trump's HHS secretary nominee, Alex Azar, will push policies that benefit pharmaceutical companies' pocketbooks?	His record suggests he’s unlikely to push pricing policies that would hit pharmaceutical companies’ pocketbooks. President Donald Trump tweeted Monday that his nominee for HHS secretary, Alex Azar will “be a star for … lower drug prices!” But the record of the former top executive for Eli Lilly, which tripled the price of a top-selling insulin drug while he led its U.S. operation, suggests a different story. Lilly is one of three drug companies targeted by a class-action lawsuit that accuses the company[...] “This nominee has impeccable big pharma credentials but non-existent credentials when it comes to concerns on drug prices,” said Peter Welch, co-chair of the House Democrats Drug Pricing Task Force.
Wikipedia	"Where is San Saba located in Rome, Italy?" Respond in 2 sentences.	San Saba is an ancient basilica church in Rome, Italy. It lies on the so-called "Piccolo Aventino", which is an area close to the ancient Aurelian Walls next to the Aventine Hill and Caelian Hill.

Dataset Description

Repository: https://github.com/akoksal/LongForm
Paper: https://arxiv.org/abs/2304.08460
Version: v1.0 - April 18, 2023
Contact: Abdullatif Köksal

License

The LongForm project is subject to a MIT License with custom limitations for restrictions imposed by OpenAI (for the instruction generation part), as well as the license of language models (OPT, LLaMA, and T5).

Citation

@misc{koksal2023longform,
      title={LongForm: Optimizing Instruction Tuning for Long Text Generation with Corpus Extraction}, 
      author={Abdullatif Köksal and Timo Schick and Anna Korhonen and Hinrich Schütze},
      year={2023},
      eprint={2304.08460},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}