🎻 Amadeus

By using Domain Adaptation on Qwen 3.5-4B, we make the model speak in the style of Amadeus, the AI that appears in the Steins;Gate 0 series.

For that, we got a dataset from the VN, and use unsloth for the fine-tuning.

Training scripts

The files in src/ are training and experimentation scripts. DAPT.py runs domain adaptation training and telegram_bot.py starts a bot for Telegram interaction.

App backend

The app-facing inference service lives in backend/inference.py, and the FastAPI server lives in backend/api.py.

TTS setup

The app uses 2 Python environments:

• .amadeus_env: main backend + Qwen 3.5 inference
• .tts_env: isolated Qwen3-TTS runtime

This split is necessary because the LLM stack and Qwen3-TTS require different transformers compatibility ranges.

The backend calls TTS through a persistent local worker running from .tts_env. That worker keeps Qwen3-TTS loaded between requests, so synthesis is much faster after the first load.

Setup:

python3.11 -m venv .amadeus_env
source .amadeus_env/bin/activate
pip install -r requirements.txt

python3.12 -m venv .tts_env
source .tts_env/bin/activate
pip install -r requirements-tts.txt

cd frontend
npm install
cd ..

Run everything at once:

./run-dev.sh

This starts:

• FastAPI backend from .amadeus_env
• persistent Qwen3-TTS worker from .tts_env
• Vite frontend from frontend/

If you want to run only the backend + TTS worker:

source .tts_env/bin/activate
python -m backend.tts_worker

source .amadeus_env/bin/activate
export AMADEUS_TTS_URL="http://127.0.0.1:8001"
python backend/api.py

It also stores its runtime data inside the project:

• .runtime/tts_home
• .runtime/mplconfig
• .runtime/amadeus_tts

You can override these locations with:

• AMADEUS_TTS_HOME
• AMADEUS_TTS_MPLCONFIGDIR
• AMADEUS_TTS_OUTPUT_DIR

Data

Inside /data there is the data used for the training + an attempt to create synthetic data, where a model was prompted to generate the questions from the answers in the dataset. Here LLMs have shown to be really bad at predicting what comes previously to a sentence, so it was discarded.

Model

The trained model is inside /qwen3.5-4b-kurisu-sg-corpus_v4

Next steps

• Create a desktop app to allow interaction with the model
• Add a React frontend and avatar layer on top of the backend API