Form MT/1 · Correspondence from a member of the public

Ministry of Tedium

Department for the Administration of Small Matters

Write to the Ministry about anything. An officer will reply in due course (in practice, a few seconds). The Ministry is fictional and its officer is a small AI model, so replies are dry, procedural and sometimes wrong. Correspondence is not stored.
0 / 500 characters
How the Ministry was made

The officer answering you is Qwen2.5-1.5B-Instruct, a small open chat model (1.5 billion parameters, Apache 2.0 licence), fine-tuned on a home graphics card to talk like a British government department.

1. Real boring text

We collected a few megabytes of genuinely dull British official writing: Commons debates and written parliamentary questions, rail accident summaries, delay reason codes, the National Rail Conditions of Travel, planning documents, public notices and listed building descriptions.

2. Example conversations

A model learns to chat from example conversations. A larger model (DeepSeek) wrote about 3,300 of them, each one shown four real excerpts from one of five departments (Parliament, planning, rail, public notices, listed buildings) so it copied their rhythm and vocabulary. Real written questions and answers from Parliament were added as they were.

3. Fine-tuning with LoRA

Rather than change all 1.5 billion numbers, LoRA freezes them and trains small extra matrices alongside: about 18 million numbers, 1.2% of the model. That took about seventeen minutes on an RTX 4070. The model keeps what it knew and only learns the voice.

Before and after

Before (instructions only)Hello! I hope you're having an exceptionally pleasant day. How may I assist you today?
After fine-tuningGood morning. The Office of Rail Apologies is operating to a reduced timetable today as a result of a minor technical delay. Your greeting has been logged under general correspondence and will be acknowledged in due course.

4. Serving it

The fine-tuned model is squeezed to 4-bit numbers (about 1 GB) and runs on an ordinary four-core server with llama.cpp. No graphics card is involved in answering you, which is why officers are occasionally all engaged.

What it gets wrong

A 1.5 billion parameter model is small. It is confidently bad at arithmetic (it once put 17 times 23 at 431), it invents books, and railway matters leak into everything. Treat its answers the way you would treat any official correspondence.