nebriq-model-classifier 🧠

A custom transformer-based classifier that predicts the semantic complexity of user prompts — trained to help intelligently route prompts to the most suitable LLM (Large Language Model).

✨ What is this?

This is the model selection brain behind Nebriq — a minimalist, AI-powered note-taking app with a strong focus on AI-assisted workflows.

Instead of hardcoding logic like:

if "analyze" in prompt:
    use("gpt-4")

…this classifier understands the prompt and determines whether it’s:

• simple: small talk, casual, or factual
• medium: instructional, contextual, or practical
• advanced: abstract, creative, analytical, or deeply technical

🧩 Why does it exist?

Users shouldn’t need to know which model is “best” for a task.

This classifier allows Nebriq to automatically route user prompts to the right LLM backend (e.g., GPT-4o-mini, Claude 3, DeepSeek, etc.) depending on semantic intent and complexity, making the UX faster, cheaper, and smarter.

🔧 How was it built?

• Trained on a small curated dataset of user-like prompts
• Built using 🤗 Transformers (distilbert-base-uncased)
• Uses the Hugging Face Trainer API
• Includes support for inference, training, and evaluation

💻 Example Inference

from transformers import pipeline

classifier = pipeline("text-classification", model="paulbg/nebriq-model-classifier")

prompt = "Model the economic effects of inflation on developing countries"
res = classifier(prompt)
print(res)
# → [{'label': 'advanced', 'score': 0.94}]

🛠️ Local Setup

poetry install
poetry run python scripts/train.py # for training
poetry run python scripts/inference.py # for inference

🤝 Use cases

• Prompt routing inside AI-native tools (like Nebriq)
• Custom moderation / filter / rerouting layers
• “AI load balancer” for hybrid LLM backends

📝 License

This project is licensed under the MIT License. See the LICENSE file for details.

🤗 View on Hugging Face