README.md

---
title: Rewrite Training Space
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Trains and upgrades morpheuslord/rewrite automatically
---

# Rewrite Training Space

Trains an upgraded version of [morpheuslord/rewrite](https://huggingface.co/morpheuslord/rewrite)
and pushes it back **only if the new model beats the previous composite score**.

## Setup

1. Go to **Settings → Secrets** in this Space
2. Add secret: `HF_TOKEN` = your HuggingFace token with **write access** to `morpheuslord/rewrite`
3. Click **Start Training** in the app

## What gets upgraded

| Parameter | Original | Upgraded |
|---|---|---|
| LoRA rank | r=8 | r=16 |
| Epochs | 5 | 10 |
| Loss function | CE only | CE + Style + Semantic |
| Effective batch | 32 | 64 |
| Evaluation | GLEU only | GLEU + BERTScore + WER composite |