metadata

YAML tags:
  - copy-paste the tags obtained with the tagging app: https://github.com/huggingface/datasets-tagging
  - WIP!

Dataset Card for P3

Table of Contents
Dataset Description
Dataset Structure
Dataset Creation
Considerations for Using the Data
Additional Information

Dataset Description

Homepage: https://bigscience.huggingface.co/promptsource
Repository: https://github.com/bigscience-workshop/promptsource/
Paper: TBA
Point of Contact: Victor Sanh (victor@huggingface.co)

Dataset Summary

P3 is a collection of prompted English datasets covering a diverse set of NLP tasks. A prompt is the combination of an input template and a target template. The templates are functions mapping a data example into natural language for the input and target sequences. For example, in the case of an NLI dataset, the data example would include fields for Premise, Hypothesis, Label. An input template would be If {Premise} is true, is it also true that {Hypothesis}?, whereas a target template can be defined with the label choices Choices[label]. Here Choices is prompt-specific metadata that consists of the options yes, maybe, no corresponding to label being entailment (0), neutral (1) or contradiction (2).

Prompts are collected using Promptsource, an interface to interactively write prompts on datasets, and collect prompt-specific metadata such as evaluation metrics. As of October 13th, there are 2'000 prompts collected for 270+ data(sub)sets. The collection of prompts is publicly available on Promptsource.

To train T0*, we used a subset of the prompts available in Promptsource (see details here). However, some of the prompts use random.choice, a method that selects uniformly at random an option in a list of valid possibilities. For reproducibility purposes, we release the collection of prompted examples used to train T0*. The data available here are the materialized version of the prompted datasets used in Multi-task enables task zero-shot generalization which represent only a subset datasets for which there is at least one prompt on Promptsource.

Supported Tasks and Leaderboards

The tasks represented in P3 cover a diverse set of NLP tasks including multiple-choice QA, sentiment analysis or natural language inference. We detail the full list of datasets in Source Data.

Languages

The data in P3 are in English (BCP-47 en).

Dataset Structure

Data Instances

An example of "train" looks as follows:

To check all the prompted examples, you can use Promptsource hosted tool and choose the Prompted dataset viewer mode in the left panel.

Data Fields

The data fields are the same among all splits:

input_text: the natural language input fed to the model
target_text: the natural language target that the model has to generate
tokenized_input: the tokenized input_text with T5's tokenizer
tokenized_target: the tokenized target_text with T5's tokenizer

Data Splits

Data(sub)set	Split	Number of examples

Dataset Creation

Curation Rationale

P3 relies on the Hugging Face Dataset library. Any public dataset in the Datasets library can be prompted. We select the datasets that have at least one subset in English and excluded datasets containing (predominantly) non-natural language examples.

We conservatively decided not to prompt datasets that contain potentially harmful content (for instance, datasets built on social media content). However, we sometimes prompt datasets that are purposefully built to measure bias and fairness of trained models, and reserve these prompted datasets (the validation or test sets) for evaluation purposes.

Source Data

Here's the full list of the datasets present in the materialized version of P3:

Multiple-Choice QA
- CommonsenseQA
- Cosmos
- DREAM
- QASC
- QUAIL
- Quarrel
- QuaRTz
- SciQ
- Social IQA
- Wiki Hop
- WiQA
- ARC
- BoolQ
- Circa
- MC-TACO
- MultiRC
- OpenBookQA
- PIQA
- RACE
Extractive QA
- Adversarial QA
- DuoRC
- Quoref
- ROPES
- TyDiQA
- CoQA
- DROP
- QA SRL
- QuAC
- ReCoRD
- SQuAD v2
Close-book QA
- Hotpot QA
- Wiki QA
- NQ Open
- Trivia QA
- Web Questions
Structure-to-text
- Common Gen
- Wiki Bio
Sentiment
- Amazon
- App Reviews
- IMDB
- Rotten Tomatoes
- Yelp
Summarization
- CNN Daily Mail
- Gigaword
- MultiNews
- SamSum
- XSum
Topic Classification
- AG News
- DBPedia
- TREC
Paraphrase Identification
- MRPC
- PAWS
- QQP
Natural Language Inference
- ANLI
- CB
- RTE
Coreference Resolution
- WSC
- Winogrande
Word Sense disambiguation
- WiC
Sentence Completion
- COPA
- HellaSwag
- Story Cloze

TODO: recheck this list to match Figure 2 in final version of paper

Annotations

The prompts available in Promptsource are collected as part of BigScience, one-year long research workshop on large multilingual models and datasets. 36 contributors affiliated with 24 institutions in 8 countries participated to the prompt collection. Contributors are in majority machine learning researchers or machine learning engineers.

The main annotation guideline was that prompts needed to be grammatical and understandable by a native English speaker with no prior experience of the tasks. Additionally, prompts that required explicit counting or numerical indexing were removed in favor of natural language variants, e.g., instead of predicting indices of a span to extract (e.g. in extractive question answering), the model was expected to copy the span's text instead. With these minimal constraints, prompt writers were encouraged to use both formal and creative prompts and various orderings of the data. Most of the prompts correspond directly to a version of the original proposed task, although we also allowed prompts that permuted the original task (for instance, generating a document from its summary) or allowed for ambiguous output (for instance, not indicating a list of available choices).

The full annotation given to the contributors can be found here. *Note to self: the link is currently being updated with the)

Licensing Information

The dataset is released under Apache 2.0.

Citation Information

WIP

Contributions

Thanks to the contributors of promptsource for adding this dataset.