Skip to content

Ground Segment Environment Variables

Remote Training Server

Variable Default Used By Description
SERVER (required) upload_to_server.sh, download_weights.sh SSH address of the remote GPU server (e.g., user@gpu-server)
ORION_SERVER_HDD_PATH (required) upload_to_server.sh Server-side directory where the dataset tarball is extracted
ORION_LOCAL_DATA_DIR (required) upload_to_server.sh Local directory to compress and upload (e.g., orion_dataset)
ORION_ARCHIVE_NAME (required) upload_to_server.sh Temporary tarball filename (e.g., orion_data.tar.gz)
ORION_REPO_DIR (required) upload_to_server.sh Server-side directory to clone/pull the ORION repo into
ORION_SERVER_WEIGHTS_PATH (required) download_weights.sh Server-side path to the orion_lora_weights directory
ORION_SERVER_REPO_PATH (required) download_weights.sh Server-side path to the full ORION repo (deleted after successful downlink)
ORION_SERVER_DATA_PATH (required) download_weights.sh Server-side path to the dataset directory (deleted after successful downlink)

Fine-Tuning

Variable Default Used By Description
ORION_DATASET_ROOT (required) fine_tune.py Directory containing orion_dataset/ (resolves both JSONL files and the relative image paths embedded inside them).

On the training server this typically matches ORION_SERVER_HDD_PATH. On a local dev machine it would be the directory containing the orion_dataset/ produced by data_gen.py (e.g., ground_segment/data).

Receiver

receiver.py uses hardcoded values and does not read environment variables. See the receiver configuration for details.

Example: Training Workflow

# Local machine: upload dataset
export SERVER=saransh@gpu-server
export ORION_SERVER_HDD_PATH=~/hdd/gaze/datasets/extras
export ORION_LOCAL_DATA_DIR=orion_dataset
export ORION_ARCHIVE_NAME=orion_data.tar.gz
export ORION_REPO_DIR=~/code/extras

cd ground_segment/data
bash upload_to_server.sh

# On the training server: fine-tune (matches ORION_SERVER_HDD_PATH from above)
export ORION_DATASET_ROOT=~/hdd/gaze/datasets/extras
cd ground_segment/training
python fine_tune.py

# Local machine: download weights after training
export ORION_SERVER_WEIGHTS_PATH=~/code/extras/ORION/ground_segment/training/orion_lora_weights
export ORION_SERVER_REPO_PATH=~/code/extras/ORION
export ORION_SERVER_DATA_PATH=~/hdd/gaze/datasets/extras/orion_dataset
cd ground_segment/training
bash download_weights.sh