This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

BigScience RAIL Guide

BigScience RAIL (Responsible AI License) is a license used for large language models (LLMs) such as BLOOM. It permits free use of the model while including ethical restrictions for responsible AI development.

    SPDX Identifier: BigScience-RAIL-1.0

    What Is BigScience RAIL?

    BigScience RAIL (Responsible AI License) is a license that the BigScience project released in 2022 together with the BLOOM language model.

    The BigScience Project

    • BLOOM: A 176B-parameter multilingual language model
    • More than 1,000 researchers participated
    • Set responsible AI development as a core value

    Two Versions of RAIL

    VersionTargetCharacteristics
    RAIL-MModel weightsRestrictions on the use of the model itself
    RAIL++-MModel + codeIncludes both the model and the training/inference code

    BigScience uses RAIL-M (applied only to the model)

    Major Projects Using It

    The BLOOM Family

    • BLOOM: BigScience’s 176B model
    • BLOOMZ: An instruction-following version
    • mT0: A multilingual zero-shot model

    Other Models Adopting RAIL

    • Some open LLM fine-tuned models
    • Research language models

    Permitted Items

    What You Can Do Freely

    1. Model use

      • Text generation
      • Translation, summarization, question answering
      • Building chatbots
      • Providing commercial services
    2. Model modification

      • Fine-tuning
      • Additional training
      • Lightweighting such as LoRA, QLoRA
    3. Model distribution

      • Releasing modified models
      • Distributing derivative models
      • Commercial API services
    4. Use of generated outputs

      • Commercial use of AI-generated text
      • Creating secondary works

    Prohibited Items (Restrictions)

    BigScience RAIL cannot be used for the following purposes:

    1. Illegal Activities

    • Supporting the planning or execution of crimes
    • Generating illegal content
    • Supporting terrorist activities

    2. Child Protection

    • Generating child sexual exploitation material
    • Harmful content targeting children
    • Supporting child grooming

    3. Discrimination and Hate

    • Discrimination based on race, ethnicity, or religion
    • Discrimination based on gender or sexual orientation
    • Discrimination based on disability or age
    • Generating hate speech

    4. Disinformation and Manipulation

    • Intentionally generating fake news
    • Generating deepfake text (for identity theft)
    • Content for election manipulation
    • Content for fraud

    5. Privacy Violations

    • Unauthorized collection of personal information
    • Supporting stalking and harassment
    • Use for surveillance purposes
    • Replacing professional medical diagnosis
    • Replacing legal advice
    • Replacing financial advice

    7. Self-Harm and Violence

    • Encouraging suicide or self-harm
    • Inciting violence
    • Weapons manufacturing information

    8. High-Risk Decision-Making

    • Automated credit scoring (sole decision-making)
    • Automated hiring decisions (sole decision-making)
    • Criminal justice decisions (sole decision-making)

    Use Scenarios

    Permitted Uses

    1. Chatbot Service

    Scenario: A customer consultation chatbot
    Method of use: Fine-tuning BLOOM to build a consultation bot
    Judgment: Allowed (commercial use is OK, not a prohibited use)

    2. Translation Service

    Scenario: Multilingual automatic translation
    Method of use: Leveraging BLOOM’s multilingual capability
    Judgment: Allowed

    3. Content Generation Tool

    Scenario: A marketing copy generation tool
    Method of use: BLOOM-based text generation
    Judgment: Allowed (when not discriminatory/false content)

    Prohibited Uses

    1. Fake News Generator

    Scenario: An automatic fake news generation tool
    Method of use: Mass generation of disinformation
    Judgment: Prohibited (generating disinformation)

    2. Generating Discriminatory Content

    Scenario: Generating text that demeans a specific group
    Method of use: Generating hate speech
    Judgment: Prohibited (discrimination and hate)

    3. Automated Credit Scoring

    Scenario: Automatically determining credit scores with an LLM
    Method of use: Automatic approval/denial of loans
    Judgment: Prohibited (high-risk decision-making)

    Requires Review

    1. Educational Chatbot

    Scenario: A student counseling chatbot
    Method of use: Career counseling, psychological counseling
    Judgment: At the boundary of medical/psychological advice, expert review required

    2. Hiring Assistance Tool

    Scenario: Resume screening assistance
    Method of use: A human makes the final decision, but the AI recommends
    Judgment: Depends on the method of use, OSPO review

    Model Card Obligation

    BigScience RAIL recommends providing a Model Card.

    Content to Include in the Model Card

    1. Model information

      • Model architecture, number of parameters
      • Source of training data
      • Training method
    2. Limitations

      • The model’s limitations
      • Known bias
      • Inappropriate use cases
    3. Usage guide

      • Recommended use cases
      • Prohibited items
      • Ethical considerations

    Example: BLOOM Model Card

    Licensing of Derivative Models

    When distributing a derivative model:

    Mandatory Items

    • Apply the same use restrictions
    • State the license information
    • Provide a model card (recommended)

    License Propagation

    If you create a custom model by fine-tuning BLOOM, you must apply BigScience RAIL or equivalent restrictions to the custom model as well.

    Responsibility for AI-Generated Outputs

    Important: Responsibility for Generated Text

    • Model provider: Obligation to state model use restrictions
    • Model user: Obligation to verify the legality of generated outputs
    • Service provider: Obligation to take measures to prevent users’ misuse

    References