README.md

metadata

annotations_creators:
  - crowdsourced
  - expert-generated
  - machine-generated
language_creators:
  - crowdsourced
  - expert-generated
  - machine-generated
  - other
languages:
  - en
licenses:
  - apache-2.0
multilinguality:
  - multilingual
  - monolingual
pretty_name: bigbench
size_categories:
  - unknown
source_datasets:
  - original
task_categories:
  - multiple-choice
  - question-answering
  - text-classification
  - text-generation
  - zero-shot-classification
  - other
task_ids:
  - multiple-choice-qa
  - extractive-qa
  - open-domain-qa
  - closed-domain-qa
  - fact-checking
  - acceptability-classification
  - intent-classification
  - multi-class-classification
  - multi-label-classification
  - text-scoring
  - hate-speech-detection
  - language-modeling

Dataset Card for BIG-bench

Table of Contents

• Table of Contents
• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Homepage/Repository: https://github.com/google/BIG-bench
• Paper: In progress
• Leaderboard:
• Point of Contact: bigbench@googlegroups.com

Dataset Summary

The Beyond the Imitation Game Benchmark (BIG-bench) is a collaborative benchmark intended to probe large language models and extrapolate their future capabilities. Tasks included in BIG-bench are summarized by keyword here, and by task name here. A paper introducing the benchmark, including evaluation results on large language models, is currently in preparation.

Supported Tasks and Leaderboards

BIG-Bench consists of both json and programmatic tasks. This implementation in HuggingFace datasets implements

• 24 BIG-bench Lite tasks

• 167 BIG-bench json tasks (includes BIG-bench Lite)

To study the remaining programmatic tasks, please see the BIG-bench GitHub repo

Languages

Although predominantly English, BIG-bench contains tasks in over 1000 written languages, as well as some synthetic and programming languages. See BIG-bench organized by keywords. Relevant keywords include multilingual, non-english, low-resource-language, translation.

For tasks specifically targeting low-resource languages, see the table below:

Task Name	Languages
Conlang Translation Problems	English, German, Finnish, Abma, Apinayé, Inapuri, Ndebele, Palauan
Kannada Riddles	Kannada
Language Identification	1000 languages
Swahili English Proverbs	Swahili
Which Wiki Edit	English, Russian, Spanish, German, French, Turkish, Japanese, Vietnamese, Chinese, Arabic, Norwegian, Tagalog

Dataset Structure

Data Instances

Each dataset contains 5 features. For example an instance from the emoji_movie task is:

{
  "idx": 0,
  "inputs": "Q: What movie does this emoji describe? 👦👓⚡️\n  choice: harry potter\n. choice: shutter island\n. choice: inglourious basterds\n. choice: die hard\n. choice: moonlight\nA:"
  "targets": ["harry potter"],
  "multiple_choice_targets":["harry potter", "shutter island", "die hard", "inglourious basterds", "moonlight"],
  "multiple_choice_scores": [1, 0, 0, 0, 0]
}

For tasks that do not have multiple choice targets, the lists are empty.

Data Fields

Every example has the following fields

• idx: an int feature
• inputs: a string feature
• targets: a sequence of string feature
• multiple_choice_targets: a sequence of string features
• multiple_choice_scores: a sequence of int features

Data Splits

Each task has a default, train and validation split. The split default uses all the samples for each task (and it's the same as all used in the bigbench.bbseqio implementation.) For standard evaluation on BIG-bench, we recommend using the default split, and the train and validation split is to be used if one wants to train a model on BIG-bench.

Dataset Creation

BIG-bench tasks were collaboratively submitted through GitHub pull requests.

Each task went through a review and meta-review process with criteria outlined in the BIG-bench repository documentation. Each task was required to describe the data source and curation methods on the task README page.

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

BIG-bench contains a wide range of tasks, some of which are sensitive and should be used with care.

Some tasks are specifically designed to test biases and failures common to large language models, and so may elicit inappropriate or harmful responses. For a more thorough discussion see the [BIG-bench paper](in progress).

To view tasks designed to probe pro-social behavior, including alignment, social, racial, gender, religious or political bias; toxicity; inclusion; and other issues please see tasks under the pro-social behavior keywords on the BIG-bench repository.

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

For a more thorough discussion of all aspects of BIG-bench including dataset creation and evaluations see the BIG-bench repository https://github.com/google/BIG-bench and paper []

Dataset Curators

[More Information Needed]

Licensing Information

Apache License 2.0

Citation Information

To be added soon !

Contributions

For a full list of contributors to the BIG-bench dataset, see the paper.

Thanks to @andersjohanandreassen and @ethansdyer for adding this dataset to HuggingFace.