What does it do?

Give the model a small modern annoyance:

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.

…and it will produce a short comic scene in the style of a badly organised Elizabethan theatre company.

Typical outputs include:

• TITLE
• DRAMATIS PERSONAE
• SCENE
• THOU MUST CHOOSE

Style

The intended voice combines influences from:

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

The goal is not historical accuracy, but readable modern dialogue with theatrical overreaction.

Training

This adapter was fine-tuned using LoRA with supervised fine-tuning (SFT) on a hand-curated dataset of approximately 240 examples.

The dataset was iteratively refined over multiple training runs to improve:

• structural consistency
• role selection
• relevance to the original grievance
• comedic escalation
• dialogue quality

On a 10-prompt held-out manual evaluation, the fine-tuned model scored 56/80, compared with 36/80 for the base Qwen2.5-3B-Instruct model using the same prompt and scoring rubric.

Design principles

The model was trained around a few simple ideas:

• Take the grievance seriously.
• Make the reaction wildly disproportionate.
• Keep the cast reasonably small and situation-specific.
• Let the comedy emerge from the original complaint.
• Never solve the problem sensibly.

Intended use

The model works best for:

• everyday annoyances
• social awkwardness
• household disasters
• transport failures
• office politics
• mildly haunted appliances
• inexplicably judgemental animals

It is intended for entertainment and creative text generation.

Limitations

This model intentionally prioritises style over factual accuracy.

Recurring characters and running jokes are expected behaviour.

The model performs best on small everyday frustrations rather than major life events.

Project structure

The project consists of three related repositories:

• LoRA adapter (this repository): The original PEFT fine-tuning.
• Merged model: A standalone Transformers version of the model.
• GGUF edition: A quantised deployment build for llama.cpp and local inference.

Quick start

This repository contains a PEFT LoRA adapter and should be loaded on top of Qwen2.5-3B-Instruct.

About the project

The objective was not to build a general-purpose assistant, but a small theatrical comedy troupe capable of treating everyday inconveniences as matters of grave national importance.

Training procedure

This model was trained with supervised fine-tuning (SFT) using TRL and PEFT.

Framework versions

• PEFT 0.19.1
• TRL 1.6.0
• Transformers 5.10.2
• PyTorch 2.11.0+cu128
• Datasets 5.0.0
• Tokenizers 0.22.2

Try the model

• 🎭 Interactive demo

Related repositories

• Merged model
• GGUF edition