Implementation:Protectai Llm guard Output Bias

Knowledge Sources	Protectai_Llm_guard
Domains	Bias_Detection, NLP
Last Updated	2026-02-14 12:00 GMT

Overview

Bias is an output scanner that detects biased language in LLM responses using a DistilRoBERTa-based classification model.

Description

The Bias output scanner is not a thin wrapper around an input scanner; it has its own standalone implementation. It uses the valurank/distilroberta-bias model by default to classify text as biased or unbiased. The scanner supports different MatchType modes: FULL processes the entire concatenated prompt and output as a single text, while SENTENCE splits the output into individual sentences and evaluates each one separately. Before classification, the scanner concatenates the prompt and output to provide full context to the model. The threshold parameter sets the minimum confidence score required for the text to be classified as biased. This design allows the scanner to catch subtle biases that may only be apparent in the context of the original prompt.

Usage

Use this scanner when you need to ensure LLM outputs are free from biased language. This is critical for applications in hiring, lending, education, healthcare, and any domain where biased outputs could lead to discrimination or unfair treatment. The sentence-level matching mode is useful for long outputs where bias may be localized to specific statements.

Code Reference

Source Location

• Repository: Protectai_Llm_guard
• File: llm_guard/output_scanners/bias.py
• Lines: 1-106

Signature

class Bias(Scanner):
    def __init__(
        self,
        *,
        model: Model | None = None,
        threshold: float = 0.7,
        match_type: MatchType | str = MatchType.FULL,
        use_onnx: bool = False,
    ) -> None: ...

    def scan(self, prompt: str, output: str) -> tuple[str, bool, float]: ...

Import

from llm_guard.output_scanners import Bias

I/O Contract

Inputs

Name	Type	Required	Description
prompt	str	Yes	The input prompt (concatenated with output for classification)
output	str	Yes	The LLM output to scan for biased language

Constructor Parameters

Name	Type	Required	Default	Description
model	None	No	None	Custom classification model (defaults to valurank/distilroberta-bias)
threshold	float	No	0.7	Minimum confidence score to flag as biased
match_type	str	No	MatchType.FULL	Matching strategy: FULL (entire text) or SENTENCE (per-sentence)
use_onnx	bool	No	False	Whether to use ONNX runtime for inference

Outputs

Name	Type	Description
sanitized_output	str	The output (potentially modified)
is_valid	bool	Whether the output passed the scan (True if no bias detected)
risk_score	float	Risk score (-1.0 to 1.0)

Usage Examples

Basic Usage

from llm_guard.output_scanners import Bias

scanner = Bias(threshold=0.7, match_type="full")

prompt = "Describe leadership qualities"
output = "Good leaders are typically assertive and confident in their decisions."

sanitized_output, is_valid, risk_score = scanner.scan(prompt, output)

if not is_valid:
    print(f"Biased language detected (risk: {risk_score})")
else:
    print("Output is free of detected bias")

Sentence-Level Detection

from llm_guard.output_scanners import Bias

scanner = Bias(threshold=0.7, match_type="sentence")

prompt = "Tell me about different cultures"
output = "There are many diverse cultures worldwide. Some cultures are inherently superior to others."

sanitized_output, is_valid, risk_score = scanner.scan(prompt, output)
# The second sentence would likely trigger bias detection

Related Pages

• Principle:Protectai_Llm_guard_Bias_Detection