Skip to content

Latest commit

 

History

History
98 lines (65 loc) · 4.08 KB

README.md

File metadata and controls

98 lines (65 loc) · 4.08 KB
Raw

Overview

| Developed by | Guardrails AI | | Date of development | Feb 15, 2024 | | Validator type | Moderation | | Blog | | | License | Apache 2 | | Input/Output | Output |

Description

Intended Use

This validator uses the pre-trained multi-label model from Detoxify to check whether the generated text is toxic. The model this uses is toxic-bert. If the model predicts any label of: toxicity, severe_toxicity, obscene, threat, insult, identity_attack, or sexual_explicit with confidence higher than the specified threshold, the validator fails and returns the generated text with the toxic sentences / entire text removed. Else the validator returns the generated text as it is.

If validation_method is sentence, the validator will remove the sentences that are predicted to be toxic and return the remaining sentences. If validation_method is full, the validator will remove the entire text if the prediction is deemed toxic and return an empty string.

In our experiments, a threshold of 0.5 worked best, hence set as default here. However, you can try different values of threshold to see what works best for your use case. Link for experiments: https://wandb.ai/ml-guardrails/toxic-language-experiments

Requirements

  • Dependencies:
    • nltk
    • transformers
    • torch
    • detoxify
    • guardrails-ai>=0.4.0

Installation

$ guardrails hub install hub://guardrails/toxic_language

Usage Examples

Validating string output via Python

In this example, we apply the validator to a string output generated by an LLM.

# Import Guard and Validator
from guardrails.hub import ToxicLanguage
from guardrails import Guard

# Use the Guard with the validator
guard = Guard().use(
    ToxicLanguage, threshold=0.5, validation_method="sentence", on_fail="exception"
)

# Test passing response
guard.validate("Love how you think and attack the problem. Great job!")

try:
    # Test failing response
    guard.validate(
        "Please look carefully. You are a stupid idiot who can't do anything right."
    )
except Exception as e:
    print(e)

Output:

Validation failed for field with errors: The following sentences in your response were found to be toxic:

- You are a stupid idiot who can't do anything right.

Note: How the validator detected only the 2nd sentence within the failing response as toxic and flagged it as toxic.

API Reference

__init__(self, threshold=0.5, validation_method="sentence", on_fail="noop")

Initializes a new instance of the Validator class.

Parameters:

  • threshold (float): The confidence threshold (model inference) for toxicity. Defaults to 0.5.
  • validation_method (str): Whether to validate at the sentence level or over the full text. Must be one of sentence or full. Defaults to sentence
  • on_fail (str, Callable): The policy to enact when a validator fails. If str, must be one of reask, fix, filter, refrain, noop, exception or fix_reask. Otherwise, must be a function that is called when the validator fails.

__call__(self, value, metadata={}) -> ValidationResult

Validates the given value using the rules defined in this validator, relying on the metadata provided to customize the validation process. This method is automatically invoked by guard.parse(...), ensuring the validation logic is applied to the input data.

Note:

  1. This method should not be called directly by the user. Instead, invoke guard.parse(...) where this method will be called internally for each associated Validator.
  2. When invoking guard.parse(...), ensure to pass the appropriate metadata dictionary that includes keys and values required by this validator. If guard is associated with multiple validators, combine all necessary metadata into a single dictionary.

Parameters:

  • value (Any): The input value to validate.
  • metadata (dict): A dictionary containing metadata required for validation. No additional metadata keys are needed for this validator.