---

license: mit title: ' Headline-booster' sdk: gradio emoji: 💻 colorFrom: yellow colorTo: yellow pinned: true sdk_version: 6.17.3

Headline Booster

Headline Booster is a small-model-ready Gradio chatbot that helps entrepreneurs create clearer, stronger headlines in seconds.

What it does

The user gives three simple pieces of information:

1. What they sell
2. Who it is for
3. The result the audience wants

The app then generates 5 persuasive headline options by default.

How to run locally

pip install -r requirements.txt
python app.py

Hackathon alignment

• Built with Gradio
• Designed for Hugging Face Spaces
• Prepared for a small model under 32B parameters
• Tiny Titan target model: Qwen/Qwen2.5-1.5B-Instruct
• Quality fallback model: Qwen/Qwen2.5-3B-Instruct
• Tiny model mode by default
• Local visual-development fallback via USE_REAL_MODEL=false
• No model selector in the public interface
• ZeroGPU integration prepared

Architecture

.
├── app.py
├── requirements.txt
├── README.md
├── frontend/
│   ├── README.md
│   └── styles.css
└── docs/
    ├── CODEX_NOTES.md
    ├── COMMIT_LOG.md
    ├── FIELD_NOTES.md
    ├── FRONTEND_REFERENCE_NOTES.md
    └── TINY_TITAN_PLAN.md

Chatbot behavior

Headline Booster now behaves like a focused chatbot instead of a static generator:

• Short greetings like hola start the conversation and ask only for the three required data points.
• The chat shows La IA está trabajando... while the model prepares the answer.
• The frontend no longer shows example buttons or a model selector, keeping the screen focused on one chat input.
• A lightweight JSON history is saved locally per Gradio session/user namespace and cleared with + Nuevo chat; no login, database, payments, or external API is required.

Performance defaults

The app uses faster defaults for the first public Space version:

• MAX_NEW_TOKENS=280 keeps the tiny model focused on 5 headline options instead of long generations.
• The old character-by-character streaming helper was removed; the app now swaps a working message for the final answer.
• Use USE_REAL_MODEL=false only when iterating on visual design locally.

Tiny Titan model mode

The app uses the tiny model path by default. For explicit Tiny Titan configuration, set:

USE_REAL_MODEL=true
MODEL_ID=Qwen/Qwen2.5-1.5B-Instruct

If the 1.5B model needs stronger copywriting quality, use the under-4B fallback:

MODEL_ID=Qwen/Qwen2.5-3B-Instruct

Source code

GitHub repo: TODO

Built with Codex

This app was built with help from OpenAI Codex as a coding agent.

Codex migration evidence is documented in docs/CODEX_NOTES.md and docs/COMMIT_LOG.md. Visual adaptation notes are documented in docs/FRONTEND_REFERENCE_NOTES.md.