Skip to content

Quickstart

Get from zero to a running training job in 5 steps. For detailed explanations of each step, see Getting Started.

How the online services connect

HuggingFace Hub                        Weights & Biases
  lerobot/pi05_base ──► pulls at start      (optional)
  your-org/dataset  ──► pulls at start          ▲
  your-org/model    ◄── pushes on finish         │ live metrics
                                                  │
                              Training run (your GPU)

HF_TOKEN authenticates all Hub transfers. WANDB_API_KEY + wandb.enable: true activates live metric streaming. Nothing is baked into the container — all credentials are passed as environment variables at runtime.

Step 1 — One-time setup

Accept the model licenses on HuggingFace (click Agree and access repository):

Upload your dataset:

huggingface-cli login --token $HF_TOKEN
huggingface-cli repo create your-dataset-name --type dataset --private
huggingface-cli upload your-hf-username/your-dataset-name /path/to/dataset/ --repo-type dataset

Step 2 — Pull the container

singularity pull vlash.sif docker://frieddeli/vlash-forge:latest

docker pull frieddeli/vlash-forge:latest

Step 3 — Set environment variables

export SCRATCH=/your/persistent/storage   # see Storage Setup for options
export HF_TOKEN=hf_xxx
export DATASET_REPO_ID=your-hf-username/your-dataset-name

Step 4 — Smoke test

./scripts/train.sh examples/train/pi05/cloud.yaml steps=100 save_freq=100 log_freq=10

Confirms dataset loads, model downloads, and a checkpoint saves. If this passes, proceed.

Step 5 — Full run

qsub scripts/train_pbs.pbs
tail -f logs/<job-id>.o

sbatch scripts/train_slurm.sbatch examples/train/pi05/cloud.yaml

cp .env.example .env   # fill in HF_TOKEN, DATASET_REPO_ID, SCRATCH
docker-compose up

Checkpoint lands at $SCRATCH/outputs/pi05_cloud/checkpoints/last/pretrained_model/. Set push_to_hub: true in cloud.yaml to upload it to HF Hub automatically.

Need more detail? See First-time Setup, Storage Setup, or Running Training.