| title | TuneOS |
|---|---|
| colorFrom | blue |
| colorTo | indigo |
| sdk | docker |
| pinned | true |
| app_port | 7860 |
TuneOS is a fine-tuning workstation for large language models. You bring a model and a dataset; TuneOS handles dataset prep, training, evaluation, and pushing the adapter to the Hub. It runs as a macOS desktop app or as two Docker containers on Hugging Face Spaces — compute stays on infrastructure you own.
You need Python 3.10+, Poetry, and Docker Desktop.
git clone https://github.com/SahilKumar75/TuneOS
cd TuneOS
cp .env.example .env # add HF_TOKEN for gated models
poetry install
docker compose up -d # starts Redis + Celery worker
poetry run reflex runOpen http://localhost:3000. For your first run, try EleutherAI/pythia-410m — it trains on CPU and finishes in a few minutes.
No Docker? Start Redis and the worker manually:
redis-server &
celery -A workers.celery_app worker --loglevel=infoThe wizard supports three paths. Pick the mode in step 1 and the rest of the form adapts.
| Mode | What it does |
|---|---|
| SFT | Supervised fine-tuning on instruction/chat data |
| DPO | Preference alignment from chosen/rejected pairs |
| KD | Knowledge distillation from a teacher model |
Six adapter techniques are available per run: LoRA, QLoRA, AdaLoRA, IA3, prefix tuning, and prompt tuning. Advanced mode lets you stack a second technique on top via PeftMixedModel.
Any Hugging Face causal LM works — target_modules are auto-detected. Well-tested defaults exist for Mistral 7B, Llama 3 8B, Phi-3 Mini, Gemma 2B, and Pythia 410M. Auto-detection also covers Qwen2/3, Falcon, StarCoder2, Mixtral, and more. See docs/supported-models.md for VRAM estimates.
- Quickstart — first fine-tune, step by step
- Architecture — how the queue, worker, and UI fit together
- Deploy to HF Spaces — two-Space setup with Upstash Redis
- API reference — all REST endpoints
poetry run pytest
poetry run ruff check .The trainer integration test (real GPU, real model) is opt-in: TUNEOS_INTEGRATION_TESTS=1 pytest tests/test_trainer_integration.py.
See CONTRIBUTING.md. Security issues go to SECURITY.md. Apache 2.0 license.