Re-usable & scalable RLHF training pipeline with Dagster and Modal.

Read full story in this blog post: RLHF with Dagster and Modal

You would need access to

• HF to save your model.
• Modal to use GPU for training.
• OpenAI to use GPT-4 for evaluation.

Make sure your .env file looks like this:

HF_TOKEN=hf_
MODAL_TOKEN_ID=ak-
MODAL_TOKEN_SECRET=as-
OPENAI_API_KEY=sk

The recommended way is to use a prebuilt Docker image.

docker pull ghcr.io/kyryl-opens-ml/rlfh-dagster-modal:main
docker run -it --env-file .env -p 3000:3000 ghcr.io/kyryl-opens-ml/rlfh-dagster-modal:main

Make sure you depliyed training & inference function to Modal.

modal deploy ./rlhf_training/serverless_functions.py

Finally run Dagster.

dagster dev -f rlhf_training/__init__.py -p 3000 -h 0.0.0.0