An end-to-end recipe generation system: a Dagster training pipeline that fine-tunes Gemma-2B on recipe data with LoRA, a FastAPI streaming server, and a React web UI (kitchen-genie).
The trained adapter is published at ClaireLee2429/gemma-2b-recipes-lora.
┌─────────────────────────────────────────────────────────────┐
│ Training Pipeline │
│ │
│ HuggingFace Hub ──► Dagster Pipeline ──► LoRA Adapter │
│ (corbt/all-recipes) (clean, tokenize, (pushed to HF Hub) │
│ split, train) │
│ │
│ Alternative: Google Colab + Unsloth (2x faster on T4 GPU) │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Inference & Serving │
│ │
│ inference.py ── CLI script for batch generation │
│ server.py ───── FastAPI + SSE streaming (POST /generate) │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Web Frontend │
│ │
│ kitchen-genie ── React + Vite + Tailwind + shadcn/ui │
│ Streams tokens from server.py in real time │
│ github.com/glee2429/kitchen-genie │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Voice Interface │
│ │
│ recipe-voice-agent ── Call a phone number, get a recipe │
│ Claude Haiku routes, Gemma-2B cooks │
│ Sends recipe via SMS after the call │
│ github.com/glee2429/recipe-voice-agent│
└─────────────────────────────────────────────────────────────┘
# Install and start the API server
pip install -e ".[serve]"
huggingface-cli login
python server.py --adapter ClaireLee2429/gemma-2b-recipes-lora
# Start the frontend (in a separate terminal)
cd ../kitchen-genie && npm install && npm run dev
# Open http://localhost:8080recipe-lm/
├── data_pipeline/ # Dagster training pipeline
│ ├── assets/ # download, clean, tokenize, split, train
│ ├── resources.py # HuggingFaceConfig resource
│ ├── io_managers.py # Arrow dataset I/O manager
│ └── definitions.py # Dagster definitions
├── notebooks/
│ └── train_unsloth.ipynb # Colab notebook (Unsloth + T4 GPU)
├── inference.py # CLI inference with post-processing
├── server.py # FastAPI streaming API server
├── configs/
│ └── default.yaml # Default pipeline configuration
├── tests/
│ └── test_assets.py # Pipeline asset tests
└── pyproject.toml # Dependencies (dev, colab, serve)
Run the Dagster pipeline locally or use the Colab notebook for free GPU training:
# Local (Dagster)
pip install -e ".[dev]"
dagster dev
# Colab (Unsloth, ~2x faster)
Pipeline: raw_dataset → cleaned_dataset → tokenized_dataset → train_val_splits → trained_model
See configs/default.yaml for all training parameters (LoRA rank, learning rate, batch size, etc.).
# CLI
python inference.py --prompt "Recipe for chocolate chip cookies:"
python inference.py --adapter ClaireLee2429/gemma-2b-recipes-lora --prompt "Recipe for soup:"
python inference.py --no-adapter --prompt "Recipe for tacos:" # base model comparison
# API server
python server.py --adapter ClaireLee2429/gemma-2b-recipes-lora
curl -N -X POST http://localhost:8000/generate \
-H "Content-Type: application/json" \
-d '{"prompt": "Recipe for pasta carbonara:"}'Run python inference.py --help or python server.py --help for all options.
pytest tests/- kitchen-genie - React web frontend for recipe generation
- recipe-voice-agent - Containerized voice interface (phone call + SMS) powered by OpenClaw and ClawdTalk