---
title: Dex Neural Bake — Tiny-Router Trainer
emoji: ⚡
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.32.1
python_version: '3.10'
app_file: app.py
pinned: false
suggested_hardware: h200
---

# ⚡ Tiny-Router Trainer (ZeroGPU / H200)

Train the multi-head routing classifier on H200 GPU, export to ONNX, and download.

## What it does

Trains [tiny-router](https://github.com/UdaraJay/tiny-router) — a compact multi-head text classifier that predicts 4 routing signals from short messages:

| Head | Labels | Purpose |
|---|---|---|
| `relation_to_previous` | new, follow_up, correction, confirmation, cancellation, closure | Is this new context or modifying prior? |
| `actionability` | none, review, act | Should we do something? |
| `retention` | ephemeral, useful, remember | How long to keep this? |
| `urgency` | low, medium, high | How fast to respond? |

## Usage

1. Click **Train on H200** with default settings (MiniLM-L12, 10 epochs, batch 32)
2. Training takes ~1-2 minutes on H200
3. Download the ONNX model when done
4. Use `onnxruntime` for local inference

## Training data

2,892 synthetic examples covering agent/product message routing scenarios.

## Credits

Based on [tiny-router](https://github.com/UdaraJay/tiny-router) by UdaraJay.