Environment Variables
All variables can be set in a .env file (copy from .env.example) or exported
in your shell before running scripts/train.sh or docker-compose up.
Required
| Variable |
Description |
SCRATCH |
Path to persistent storage on the host, bind-mounted to /scratch inside the container. See Storage Setup. |
HF_TOKEN |
HuggingFace token with read+write access. Required to download gated models (π₀/π₀.₅) and push trained checkpoints. |
DATASET_REPO_ID |
HuggingFace dataset repo ID, e.g. your-org/your-dataset. Substituted into dataset.repo_id in the training config at runtime. |
Optional
| Variable |
Default |
Description |
NUM_GPUS |
1 |
Number of GPUs to use. Passed to accelerate launch --num_processes. |
TRAIN_BACKEND |
deepspeed |
Training backend. deepspeed for LoRA, fsdp for full fine-tuning. |
WANDB_API_KEY |
— |
W&B API key. Only needed when wandb.enable: true in your training config. |
Inside the container
These are set automatically by the entrypoint and do not need to be set manually:
| Variable |
Value |
Description |
HF_HOME |
/scratch/.cache/huggingface |
HuggingFace cache directory. Persists via the /scratch bind-mount. |
Example .env
# Copy .env.example to .env and fill in:
HF_TOKEN=hf_xxx
DATASET_REPO_ID=your-org/your-dataset
SCRATCH=/your/persistent/storage
NUM_GPUS=4
TRAIN_BACKEND=deepspeed
WANDB_API_KEY=