build-small-hackathon/tragedy-of-the-group-chat

🎭 The Tragedy of the Group Chat

Most AI assistants try to solve your problems.

This one assembles a badly organised Elizabethan theatre company and turns your minor inconvenience into a short comedy.

You forgot to buy milk. Again. Everyone is angry. A sensible person would apologise. A sensible model would offer practical advice. The Tragedy of the Group Chat instead assembles an Elizabethan comedy troupe to make the situation considerably worse

What does it do?

Enter a tiny modern grievance.

The model will produce a short theatrical catastrophe featuring:

• —TITLE
• —DRAMATIS PERSONAE
• —SCENE
• —THOU MUST CHOOSE

Inspired by:

• —Shakespeare
• —Blackadder
• —Monty Python
• —British sitcoms
• —Amateur dramatic societies

The goal is not historical accuracy.

The goal is to make a completely ordinary problem feel absurdly important.

Demo video: youtube

Try these examples

• —My neighbour's cat judged me through the window.
• —The bakery sold out of the exact pastry I had emotionally chosen.
• —The automatic lights turned off while I was still in the shower.
• —Someone took the last trolley with a working wheel.
• —I said good morning to someone and they just blinked.
• —The bus drove straight past me.

About the project

The Tragedy of the Group Chat was created for the Hugging Face Build-Small Hackathon (June 2026) - Track 2: An Adventure in Thousand Token Wood.

The project explores how much personality and structure can be taught to a relatively small local language model using a compact hand-written dataset, iterative fine-tuning and efficient local deployment.

The model was:

• —fine-tuned with LoRA,
• —merged into a 3B parameter base model,
• —converted to GGUF,
• —quantised for efficient local inference,
• —packaged for both Transformers and llama.cpp deployments,
• —and deployed as a custom Gradio application.

The interactive Space uses the merged Transformers model for responsive inference, while the project also includes a standalone GGUF deployment build.

The entire experience runs locally. No cloud language model APIs are used.

Performance Evaluation

A held-out manual benchmark of ten prompts was used throughout development.

Base Qwen2.5-3B-Instruct: 36/80

The Tragedy of the Group Chat: 56/80

The benchmark evaluated:

• —structural consistency,
• —relevance,
• —comedic voice,
• —and deployment readiness.

The fine-tuned model outperformed the base model against these benchmarks.

Project repositories

• —🎭 LoRA adapter
• —🎭 Merged model
• —🎭 GGUF edition

Built with

• —Qwen2.5-3B-Instruct
• —LoRA fine-tuning
• —GGUF
• —Gradio
• —Hugging Face Spaces

Demo video: youtube

Social Media post LinkedIn

Created by: sdavies

Philosophy

Take the grievance seriously.

Make the reaction wildly disproportionate.

Never solve the problem sensibly.