🚀 Introducing HallOumi - A State-of-the-Art Claim Verification Model

Introducing HallOumi, a state-of-the-art claim verification model, outperforming DeepSeek R1, OpenAI o1, Google Gemini 1.5 Pro, and Anthropic Sonnet 3.5 at only 8 billion parameters!

HallOumi, the hallucination detection model built with Oumi, is a system built specifically to enable per-sentence verification of any content (either AI or human-generated) with sentence-level citations and human-readable explanations.

Try a hosted version of this demo on our website!

Read more in our blog post here!

⚡ Quickstart

🐳 Docker build

You can easily build and run the HallOumi demo via docker. While inside the repo directory:

docker build -t halloumi-demo .
docker run -p 3000:3000 halloumi-demo

Open http://localhost:3000 with your browser to access the demo!

💻 Building from source

Dev Environment Setup

1. Install NPM

   brew install npm
   

2. NextJS and React

   npm install next@latest react@latest react-dom@latest
   

Running Locally

First, run the development server:

npm run dev
# or
yarn dev
# or
pnpm dev
# or
bun dev

Open http://localhost:3000 with your browser to access the demo!

🤖 Self-host Models

If you'd like to point the demo to your own self-hosted version of the models, simply modify data.json.

Below is a sample data.json assuming you've self hosted the generative model at localhost:8000:

{
    "models": [
        {
            "displayName": "My custom hosted model",
            "name": "mymodel",
            "apiUrl": "http://localhost:8000/chat/completions",
            "isEmbeddingModel": false
        }
    ],
    "examples": [
        {
            "displayName": "Getting Started",
            "claim": "Text here appears in the 'Claims to verify' box.",
            "context": "This text will appear in the 'Context' box."
        }
    ]
}

The demo assumes that the target endpoint supports the standard OpenAI API for each model.

Host the Generative HallOumi model

pip install sglang
python3 -m sglang.launch_server --model-path oumi-ai/HallOumi-8B --port 8000 --dtype auto --mem-fraction-static 0.9 --trust-remote-code

Host the Classifier HallOumi model

Note that the classifier is hosted as an embeddings model. You can query it via the standard /embeddings endpoint in OpenAI-compatible API servers.

pip install sglang
python3 -m sglang.launch_server --model-path oumi-ai/HallOumi-8B-classifier --port 8001 --dtype auto --mem-fraction-static 0.9 --trust-remote-code --is-embedding