README.md

metadata

title: G0 Hallucination Detector
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
short_description: Detect when LLMs hallucinate using 3-criterion grounding

G0 Hallucination Detector

Detect when LLMs make things up using a 3-criterion grounding metric.

How It Works

G0 = (Tracking × Intervention × Counterfactual)^(1/3)

• Tracking: Does the claim semantically follow from the sources?
• Intervention: Would changing the sources change the claim?
• Counterfactual: Is the claim uniquely dependent on these sources?

Scores

• 0.7-1.0: Grounded - claim is well-supported
• 0.4-0.7: Partial - some support, may contain unsupported elements
• 0.0-0.4: Hallucination - claim not supported by sources

Use Cases

• Verify LLM outputs before production
• Audit RAG pipeline responses
• Research on hallucination detection

API

import gradio_client

client = gradio_client.Client("crystalline-labs/g0-detector")
result = client.predict(
    claim="The Eiffel Tower was built in 1889",
    sources="The Eiffel Tower was constructed from 1887 to 1889.",
    api_name="/predict"
)

Built by Crystalline Labs