---
title: PhD Research OS - Model Training
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: "5.32.0"
app_file: app.py
pinned: false
suggested_hardware: zero-a10g
---

# 🧠 PhD Research OS — ZeroGPU Training

Train the Research OS brain (Qwen2.5-3B-Instruct + QLoRA) using free ZeroGPU H200.

## Setup

1. Go to Space Settings → Hardware → Select **ZeroGPU**
2. Click "Train Micro-Batch" — each click trains ~20 steps in ~55 seconds
3. Click multiple times to accumulate training
4. Test with the inference tab
5. Push to Hub when satisfied

## How It Works

ZeroGPU gives 60-second GPU bursts. This Space trains in micro-batches:
- Each click = 20 gradient steps on H200 GPU
- Checkpoint saved after each batch
- Next click resumes from checkpoint
- ~200 total steps recommended (click ~10 times)