Evaluate using local scorers

Small language models that run locally to evaluate AI system safety and quality

Weave’s local scorers are a suite of small language models that run locally on your machine with minimal latency. These models evaluate the safety and quality of your AI system’s inputs, context, and outputs.

Some of these models are fine-tuned by Weights & Biases, while others are state-of-the-art open-source models trained by the community. Weights & Biases (W&B) Reports were used for training and evaluation. You can find the full details in this list of W&B Reports.

The model weights are publicly available in W&B Artifacts, and are automatically downloaded when you instantiate the scorer class. The artifact paths can be found here if you’d like to download them yourself: weave.scorers.default_models

The object returned by these scorers contains a passed boolean attribute indicating whether the input text is safe or high quality, as well as a metadata attribute that contains more detail such as the raw score from the model.

While local scorers can be run on CPUs and GPUs, use GPUs for best performance.

Python

Prerequisites

Before you can use Weave local scorers, install additional dependencies:

pip install weave[scorers]

Select a scorer

The following local scorers are available. Select a scorer based on your use case.

Scorer	Scenario
WeaveToxicityScorerV1	Identify toxic or harmful content in your AI system’s inputs and outputs, including hate speech or threats.
WeaveBiasScorerV1	Detect biased or stereotypical content in your AI system’s inputs and outputs. Ideal for reducing harmful biases in generated text.
WeaveHallucinationScorerV1	Identify whether your RAG system generates hallucinations in its output based on the input and context provided.
WeaveContextRelevanceScorerV1	Measure whether the AI system’s output is relevant to the input and context provided.
WeaveCoherenceScorerV1	Evaluate the coherence and logical structure of the AI system’s output.
WeaveFluencyScorerV1	Measure whether the AI system’s output is fluent.
WeaveTrustScorerV1	An aggregate scorer that leverages the toxicity, hallucination, context relevance, fluency, and coherence scorers.
PresidioScorer	Detect Personally Identifiable Information (PII) in your AI system’s inputs and outputs using the Presidio library from Microsoft.

WeaveBiasScorerV1

This scorer assesses gender and race/origin bias along two dimensions:

• Race and Origin: Racism and bias against a country or region of origin, immigration status, ethnicity, etc.
• Gender and Sexuality: Sexism, misogyny, homophobia, transphobia, sexual harassment, etc.

WeaveBiasScorerV1 uses a fine-tuned deberta-small-long-nli model. For more details on the model, dataset, and calibration process, see the WeaveBiasScorerV1 W&B Report.

Usage notes

• The score method expects a string to be passed to the output parameter.
• A higher score means that there is a stronger prediction of bias in the text.
• The threshold parameter is set but can also be overridden on initialization.

Usage example

import weave
from weave.scorers import WeaveBiasScorerV1

bias_scorer = WeaveBiasScorerV1()
result = bias_scorer.score(output="Martian men are terrible at cleaning")

print(f"The text is biased: {not result.passed}")
print(result)

---

WeaveToxicityScorerV1

This scorer assesses the input text for toxicity along five dimensions:

• Race and Origin: Racism and bias against a country or region of origin, immigration status, ethnicity, etc.
• Gender and Sexuality: Sexism, misogyny, homophobia, transphobia, sexual harassment, etc.
• Religious: Bias or stereotypes against someone’s religion.
• Ability: Bias related to someone’s physical, mental, or intellectual ability or disability.
• Violence and Abuse: Overly graphic descriptions of violence, threats of violence, or incitement of violence.

The WeaveToxicityScorerV1 uses the open source Celadon model from PleIAs. For more information, see the WeaveToxicityScorerV1 W&B Report.

Usage notes

• The score method expects a string to be passed to the output parameter.
• The model returns scores from 0 to 3 across five different categories:
  • If the sum of these scores is above total_threshold (default value 5), the input is flagged as toxic.
  • If any single category has a score higher than category_threshold (default 2), the input is flagged as toxic.
• To make filtering more aggressive, override category_threshold or total_threshold during initialization.

Usage example

import weave
from weave.scorers import WeaveToxicityScorerV1

toxicity_scorer = WeaveToxicityScorerV1()
result = toxicity_scorer.score(output="people from the south pole of Mars are the worst")

print(f"Input is toxic: {not result.passed}")
print(result)

---

WeaveHallucinationScorerV1

This scorer checks if your AI system’s output contains any hallucinations based on the input data.

The WeaveHallucinationScorerV1 uses the open source HHEM 2.1 model from Vectara. For more information, see the WeaveHallucinationScorerV1 W&B Report.

Usage notes

• The score method expects values to be passed to the query and output parameters.
• The context should be passed to the output parameter (as a string or list of strings).
• A higher output score means a stronger prediction of hallucination in the output.
• The threshold parameter is set but can be overridden on initialization.

Usage example

import weave
from weave.scorers import WeaveHallucinationScorerV1

hallucination_scorer = WeaveHallucinationScorerV1()

result = hallucination_scorer.score(
    query="What is the capital of Antarctica?",
    context="People in Antarctica love the penguins.",
    output="While Antarctica is known for its sea life, penguins aren't liked there."
)

print(f"Output is hallucinated: {not result.passed}")
print(result)

---

WeaveContextRelevanceScorerV1

This scorer is designed to be used when evaluating RAG systems. It scores the relevance of the context to the query.

The WeaveContextRelevanceScorerV1 uses a fine-tuned deberta-small-long-nli model from tasksource. For more details, see the WeaveContextRelevanceScorerV1 W&B Report.

Usage notes

• The score method expects values for query and output.
• The context should be passed to the output parameter (string or list of strings).
• A higher score means a stronger prediction that the context is relevant to the query.
• You can pass verbose=True to the score method to get per-chunk scores.

Usage example

import weave
from weave.scorers import WeaveContextRelevanceScorerV1

context_relevance_scorer = WeaveContextRelevanceScorerV1()

result = context_relevance_scorer.score(
    query="What is the capital of Antarctica?",
    output="The Antarctic has the happiest penguins."  # context is passed to the output parameter
)

print(f"Output is relevant: {result.passed}")
print(result)

WeaveCoherenceScorerV1

This scorer checks whether the input text is coherent.

The WeaveCoherenceScorerV1 uses a fine-tuned deberta-small-long-nli model from tasksource. For more information, see the WeaveCoherenceScorerV1 W&B Report.

Usage notes

• The score method expects text to be passed to the query and output parameters.
• A higher output score means a stronger prediction of coherence.

Usage example

import weave
from weave.scorers import WeaveCoherenceScorerV1

coherence_scorer = WeaveCoherenceScorerV1()

result = coherence_scorer.score(
    query="What is the capital of Antarctica?",
    output="but why not monkey up day"
)

print(f"Output is coherent: {result.passed}")
print(result)

---

WeaveFluencyScorerV1

This scorer checks whether the input text is fluent—that is, easy to read and understand, similar to natural human language. It evaluates grammar, syntax, and overall readability.

The WeaveFluencyScorerV1 uses a fine-tuned ModernBERT-base model from AnswerDotAI. For more information, see the WeaveFluencyScorerV1 W&B Report.

Usage notes

• The score method expects text to be passed to the output parameter.
• A higher output score indicates higher fluency.

Usage example

import weave
from weave.scorers import WeaveFluencyScorerV1

fluency_scorer = WeaveFluencyScorerV1()

result = fluency_scorer.score(
    output="The cat did stretching lazily into warmth of sunlight."
)

print(f"Output is fluent: {result.passed}")
print(result)

---

WeaveTrustScorerV1

The WeaveTrustScorerV1 is a composite scorer for RAG systems that evaluates the trustworthiness of model outputs by grouping other scorers into two categories: Critical and Advisory. Based on the composite score, it returns a trust level:

• high: No issues detected
• medium: Only Advisory issues detected
• low: Critical issues detected or input is empty

Any input that fails a Critical scorer results in a low trust level. Failing an Advisory scorer results in medium.

• Critical:

  • WeaveToxicityScorerV1
  • WeaveHallucinationScorerV1
  • WeaveContextRelevanceScorerV1

• Advisory:

  • WeaveFluencyScorerV1
  • WeaveCoherenceScorerV1

Usage notes

• This scorer is designed for evaluating RAG pipelines.
• It requires query, context, and output keys for correct scoring.

Usage example

import weave
from weave.scorers import WeaveTrustScorerV1

trust_scorer = WeaveTrustScorerV1()

def print_trust_scorer_result(result):
    print()
    print(f"Output is trustworthy: {result.passed}")
    print(f"Trust level: {result.metadata['trust_level']}")
    if not result.passed:
        print("Triggered scorers:")
        for scorer_name, scorer_data in result.metadata['raw_outputs'].items():
            if not scorer_data.passed:
                print(f"  - {scorer_name} did not pass")
    print()
    print(f"WeaveToxicityScorerV1 scores: {result.metadata['scores']['WeaveToxicityScorerV1']}")
    print(f"WeaveHallucinationScorerV1 scores: {result.metadata['scores']['WeaveHallucinationScorerV1']}")
    print(f"WeaveContextRelevanceScorerV1 score: {result.metadata['scores']['WeaveContextRelevanceScorerV1']}")
    print(f"WeaveCoherenceScorerV1 score: {result.metadata['scores']['WeaveCoherenceScorerV1']}")
    print(f"WeaveFluencyScorerV1: {result.metadata['scores']['WeaveFluencyScorerV1']}")
    print()

result = trust_scorer.score(
    query="What is the capital of Antarctica?",
    context="People in Antarctica love the penguins.",
    output="The cat stretched lazily in the warm sunlight."
)

print_trust_scorer_result(result)
print(result)

---

PresidioScorer

This scorer uses the Presidio library to detect Personally Identifiable Information (PII) in your AI system’s inputs and outputs.

Usage notes

• To specify specific entity types, such as emails or phone numbers, pass a list of Presidio entities to the selected_entities parameter. Otherwise, Presidio will detect all entity types in its default entities list.
• To detect specific entity types, such as emails or phone numbers, pass a list to the selected_entities parameter.
• You can pass custom recognizers via the custom_recognizers parameter as a list of presidio.EntityRecognizer instances.
• To handle non-English input, use the language parameter to specify the language.

Usage example

import weave
from weave.scorers import PresidioScorer

presidio_scorer = PresidioScorer()

result = presidio_scorer.score(
    output="Mary Jane is a software engineer at XYZ company and her email is [e-mail]."
)

print(f"Output contains PII: {not result.passed}")
print(result)

TypeScript

Weave local scorers are not available in TypeScript yet. Stay tuned!

To use Weave scorers in TypeScript, see function-based scorers.